DEV Community: JXIONG

The 2 P.M. Caffeine Rule: Your Afternoon Coffee Is Still in Your Blood at Midnight

JXIONG — Thu, 22 Oct 2026 11:47:16 +0000

It's 4 p.m., you're fading, and you get an Americano to push through the last stretch. At midnight you're in bed, tired but wired, staring at the ceiling.

Or maybe you're the other type: "Coffee doesn't affect me, I fall asleep instantly." And yet somehow you're still short on sleep every day.

Both of these are the same story, and it's a chemistry story.

The half-life account

While you're awake, your brain keeps building adenosine — the sleepiness signal. Enough accumulates, you feel sleepy. Caffeine looks almost exactly like adenosine. It sits in adenosine's receptors without passing the signal on, like taping over a doorbell button: the sleepiness is still there, you just can't hear the bell.

The problem is how slowly it leaves. Caffeine's half-life is about 5 hours (individual range roughly 2–10). So:

Cup at 2 p.m. → about half still circulating at 7 p.m.
→ about a quarter still there at midnight.

A quarter of a strong coffee at bedtime is not nothing. It's a pharmacological dose sitting on your doorbell.

The experiments are unforgiving

Drake 2013 gave participants 400 mg of caffeine (about two large Americanos) at 0, 3, or 6 hours before bed, and measured sleep with instruments rather than self-report. Even the 6-hours-before group lost more than an hour of total sleep — while the participants themselves reported noticing much less. Caveat: only 12 people, and a high dose.

Gardiner 2023 pooled 24 studies: caffeine cut total sleep by about 45 minutes on average. Their estimate for a normal cup (~100 mg): finish it at least 8.8 hours before bed. Sleep at 11 p.m.? That's a 2 p.m. cutoff.

This is why it's a rule and not a suggestion. On this one point, how you feel cannot be trusted.

Why "it doesn't affect me" is unreliable

Caffeine doesn't have to keep you awake to cost you. It changes the structure of sleep: slower to fall asleep, more light sleep, more awakenings during the night. People usually don't notice any of it — they wake up having lost an hour and rate their night as "fine."

There's real individual variation, and it's genetic. Caffeine is broken down mainly by a liver enzyme called CYP1A2, whose activity depends on your genes, smoking, oral contraceptives, pregnancy, and more. That's why one colleague drinks espresso at 9 p.m. and sleeps like a stone, and you don't.

What to do

No caffeine after 2 p.m. (adjust earlier if you're a slow metabolizer or a light sleeper)
Audit the hidden sources: strong tea, milk tea, cola, energy drinks, and some painkillers and cold medicines.
Afternoon lift without caffeine: walk around the block, get a few minutes of sun, or switch to decaf. The walk is doing more for your alertness than the coffee was.

For programmers: a late coffee to hit a deadline is trading tomorrow's performance for tonight's alertness — and sleep debt compounds, which is the next entry in this series.

Honesty corner: how strong is this?

Grade B. A meta-analysis of 24 studies plus a controlled trial, both measuring sleep objectively. Why not A: the outcomes are sleep measures, not long-term health outcomes, and the key trial had a small sample and a high dose.

Limits worth stating: caffeine clears much more slowly during pregnancy — follow your doctor's advice on intake. People with heart rhythm problems or anxiety disorders are more sensitive and should be more conservative.

This is entry #4 of a series from a free, open-source, evidence-graded guide — 111 studies, every claim graded A–D, every action tagged do it / try it / know it / don't spend:

GitHub (bilingual): DeepoMe/Awesome-Programmer-Longevity
Mobile ebook: deepkang.com/ebook/ebook.html

Previous: Sitting Is Its Own Risk Factor, and Your Gym Hour Does Not Cancel It. Next: sleep debt is compound interest — why you can't repay a week of short nights with one long lie-in.

Health education, not medical advice. If you have a medical condition or take medication, consult a physician before changing anything.

Sitting Is Its Own Risk Factor, and Your Gym Hour Does Not Cancel It

JXIONG — Thu, 22 Oct 2026 11:32:04 +0000

You sat down to fix one bug. You looked up and three hours were gone. Your back is stiff, your legs are numb, your head is foggy. So in the evening you put in a solid hour at the gym and tell yourself: that cancels today out, right?

It doesn't. Not the way you think.

This is the single lowest-tech item in a guide that reads 111 PubMed studies, and possibly the highest-leverage one. Here is what the data actually says, and it's cleaner than most fitness advice.

Sitting costs you on two separate accounts

Population research separates them, and that separation is the whole point.

Account one — total time. A harmonized meta-analysis pooling data from more than a million people found that more hours sitting per day meant higher mortality risk. The dose that offset it: roughly 60–75 minutes a day of moderate-to-vigorous activity. So yes, exercise pays something back — but you need a lot of it, and it's a payment, not a prevention.

Account two — unbroken stretches. A national cohort of nearly 8,000 Americans aged 45+, measured objectively with wearable trackers rather than self-report, found that both more total sitting and longer uninterrupted bouts were linked to higher mortality. People high on both — sat a lot, and often for 10+ minutes without moving — had the highest risk.

The honest caveat: the two accounts are hard to fully separate statistically, and how much "bout length" matters on its own is still unsettled. But the direction is consistent across both.

So an hour at the gym is great. It just doesn't pay off three unbroken hours in the chair.

Why muscle goes into standby

Think of skeletal muscle as the body's biggest recycling center for blood sugar and blood fats. Sit still and the recycling center drops into standby: the sugar and fat in your blood queue up with nobody processing them.

The mechanism is measurable. Animal studies found that during prolonged inactivity, lipoprotein lipase (LPL) — the enzyme in skeletal muscle that breaks down blood fats — drops sharply. And in human experiments, breaking up sitting with short bouts of activity every 20–30 minutes lowered post-meal blood sugar and insulin (Dunstan 2012). Contracting muscle pulls glucose into cells on its own, without waiting for insulin to unlock the door.

Stand up and walk a few steps, and the recycling center comes back online. That's it. That's the intervention.

The 2-minute rule

Set a silent repeating reminder every 30 minutes — phone, watch, Pomodoro timer, whatever.
When it fires, stand for 2–3 minutes: water, bathroom, five squats, walk to the window and focus on something far away.
It does not need to be intense. What matters is breaking up the sitting.

For programmers specifically: your build, your test run, your CI pipeline is a built-in cue. Every waiting period is a standing period. You already have dozens of them a day — you just spend them in the chair.

A surprise on the back question

Since we're on the subject of desks: the classic spinal disc pressure measurements found sitting put about 40% more pressure on the discs than standing, and leaning forward while sitting nearly doubled it (Nachemson 1981).

Then in 1999, a re-test with more precise sensors found something different: relaxed sitting is about the same as standing. Only leaning forward raises pressure significantly (Wilke 1999).

Which flips the advice. The thing to avoid is not sitting — it's the forward hunch you do while sitting. A thousand-dollar standing desk that you lower all the way down is not the fix. Sitting back, with support, is closer to neutral than perching forward over a laptop.

Honesty corner: how strong is this?

Grade B. Two large observational studies pointing the same way (Ekelund 2016; Diaz 2017), plus a short-term human experiment on post-meal glucose (Dunstan 2012). Why not A: these are observational; nobody has run a long-term randomized trial of "stand up every 30 minutes → live longer."

Two limits worth stating plainly:

Gentle movement is enough. This is not a workout program.
If you get dizzy standing up (orthostatic hypotension), pause a few seconds at the edge of the chair before rising.

This is entry #3 of a series from a free, open-source, evidence-graded guide — 111 studies, every claim graded A–D, every action tagged do it / try it / know it / don't spend:

GitHub (bilingual): DeepoMe/Awesome-Programmer-Longevity
Mobile ebook: deepkang.com/ebook/ebook.html

Previous: TCM Says “Yang Qi” Warms, Propels, and Protects. Your Mitochondria Do Exactly Those Three Things.. Next: the 2 P.M. caffeine rule — why your afternoon coffee is still in your blood at midnight.

Health education, not medical advice. If you have a medical condition or take medication, consult a physician before changing anything.

TCM Says “Yang Qi” Warms, Propels, and Protects. Your Mitochondria Do Exactly Those Three Things.

JXIONG — Thu, 22 Oct 2026 02:38:01 +0000

Traditional Chinese Medicine describes yang-qi with three verbs: it warms the body (温煦), propels it (推动), and protects it (固护). Weak yang-qi, in the classical picture, is the person who is always cold, low on energy, and run down — a description TCM has used for two thousand years.

Now look at what mitochondria do: thermogenesis (brown-fat mitochondria literally convert fuel into heat), energy conversion (every ATP you spend thinking), and quality control (mitophagy, recycling damaged units).

Three-for-three. And in 2022, a Chinese team formally proposed exactly this mapping in the academic literature: mitochondria as the physical substrate of "yang-qi" (Luo 2022).

This is where most TCM-meets-science essays wave their hands and stop. We did the opposite — the guide treats the mapping as what it is: a hypothesis, graded D (tradition + mechanism analogy, no direct test). And then asks the only interesting question: if the mapping were true, what would follow from it?

What follows: a warming-food list the classics never had

If "warming foods" = foods that help mitochondria, then the food list should include things no classical TCM text would ever call warming — graded on their own human evidence, not on tradition:

Pomegranate & walnuts → urolithin A. Activates mitophagy — the recycling of worn-out mitochondria. Positive human RCTs as supplements (muscle strength in middle-aged adults; the older-adults trial improved some endurance measures but missed its primary walking-distance endpoint); food conversion depends on your gut flora. B (supplement trials).
Beets & spinach → dietary nitrate. Makes human energy production measurably more fuel-efficient (Larsen 2011). B.
Mushrooms → ergothioneine. A longevity vitamin with a dedicated transporter in your cells; a Japanese cohort (1,344 older adults, 11+ years) found higher blood levels associated with lower dementia risk. Observational only. C.

None of these are "warm" foods by TCM classification. All of them do mitochondrial work. That's the point of a mapping with teeth — it generates predictions that tradition alone would never make.

Even chili peppers fit: capsinate activates brown-fat thermogenesis via TRPV1, with human studies behind it (B). The classics got that one right by observation; the mapping explains why.

The falsifiable prediction

Here's what makes this more than vibes. If yang-qi ≈ mitochondria, then people who self-report "yang deficiency" (chronically cold, fatigued) should show lower brown-fat activity on a PET-CT scan. That's a standard, existing clinical measurement.

Run the scan. If the correlation isn't there, the mapping dies. A hypothesis that can die is a hypothesis worth having.

For programmers, the whole thing compresses to three lines:

train your mitochondria   // Zone 2, the cheapest lever
repair your mitochondria  // sleep
feed your mitochondria    // the new warming-food list

Honesty corner: how strong is this?

The mapping itself: D. Nobody has run the PET-CT study. It is an analogy with a literature pedigree, not a finding.

The foods: graded B–C individually, on their own evidence — the list stands even if the mapping falls. And one border you must respect before touching any of this: chronic coldness and fatigue are also the classic presentation of hypothyroidism and anemia. Rule those out with a doctor first — the guide says so explicitly.

That's the two-ruler way: the mapping is labeled D, the foods are labeled B–C, and neither borrows credibility from the other.

This is entry #2 of a series from a free, open-source, evidence-graded guide — 111 studies, every claim graded A–D, every action tagged do it / try it / know it / don't spend:

GitHub (bilingual): DeepoMe/Awesome-Programmer-Longevity
Mobile ebook: deepkang.com/ebook/ebook.html

Previous: Your Brain Breaks Its Own DNA to Learn. Next: why sitting is its own risk factor, even if you exercise.

Health education, not medical advice. If you have a medical condition or take medication, consult a physician before changing anything.

Your Brain Breaks Its Own DNA to Learn — Literally

JXIONG — Wed, 21 Oct 2026 00:50:36 +0000

To write down a new memory, a neuron deliberately cuts its own DNA.

Not metaphorically. Double-strand breaks, at specific spots, on purpose — and the neurons that survive the cut-and-repair cycle are the ones that get recruited into the memory circuit.

If learning a new framework ever felt like your brain was an overloaded server mid-rebuild, that's because it was one.

The evidence chain (all top journals, all animal models)

Step 1 — The cut (Cell, 2015). When a neuron fires strongly, an enzyme called topoisomerase IIβ deliberately snips DNA near certain genes. The cut lets "immediate-early genes" switch on fast — the first molecular step of writing a memory.

Step 2 — The repair crew (Nature, 2023). Synaptic activity recruits a protein complex called NPAS4–NuA4, which physically ties neuronal activity to DNA repair. The cell that broke itself now has a dedicated team stitching it back.

Step 3 — Break survivors join the circuit (Nature, 2024). Hours after mice learn a task, a small cluster of hippocampal neurons shows persistent DNA double-strand breaks, and a "DNA-sensing" repair pathway (TLR9) kicks in. Shut that pathway down, and the mice's memory gets worse. The neurons that went through break-and-repair are exactly the ones absorbed into the memory trace.

Three papers, one arc: controlled damage → dedicated repair → damage-experienced neurons become memory.

What this means for programmers

Learning has a physical cost, and it's paid in genome integrity. Your brain isn't storing that new framework in some clean abstraction layer — it's restructuring hardware, with breaks and patches, in the neurons doing the learning.

Two practical consequences:

Compile time is real. Memory consolidation happens most efficiently during sleep. Cramming a new stack all night and going straight to a deadline means the "build" never finishes — you get the syntax but not the fluency. Give the repair crew a night.
"It hurts" is not a stop signal. Feeling fried while learning something new is consistent with what's happening at the molecular level. The right response is recovery (sleep, breaks), not abandonment — and not ignoring it either.

Honesty corner: how strong is this evidence?

Grade C — mechanism and preliminary. Top-journal chain, but the direct experiments are in mice and cells. Nobody has watched this happen in a living human brain (you'd have to watch single-neuron DNA in action, which we can't do yet).

That's why the action tier in the guide is 📖 know it — not 🧪 try it, not ✅ do it. The claim changes how you interpret the experience of learning; it doesn't hand you a new supplement.

This is entry #1 of a series. The source is a free, open-source, evidence-graded guide — 111 studies, every claim carries an evidence grade (A–D) and an action tier (do it / try it / know it / don't spend):

GitHub (bilingual): DeepoMe/Awesome-Programmer-Longevity
Mobile ebook: deepkang.com/ebook/ebook.html

Next in the series: why sitting is its own risk factor, even if you exercise.

Health education, not medical advice. If you have a medical condition or take medication, consult a physician before changing anything.

I Read 111 Longevity Studies So Programmers Don't Have To

JXIONG — Sun, 18 Oct 2026 05:16:20 +0000

Programmers joke about "surviving on coffee." I took the joke seriously: over the past months I read through decades of sleep medicine, nutritional epidemiology, and neuroscience research — 111 studies, every one verifiable on PubMed — and turned them into a free, open-source, bilingual (EN/中文) guide for developers.

Repo: DeepoMe/Awesome-Programmer-Longevity · Mobile ebook: deepkang.com/ebook/ebook.html

The one design decision that matters

Most health content gives every claim the same weight — "drink more water" and "take this compound" delivered in the same confident tone. This guide refuses to do that. Every single entry carries two ratings:

Ruler 1 — Evidence grade:

A — hard human evidence (RCTs, large cohorts, meta-analyses)
B — human evidence (observational, smaller samples)
C — mechanism & preliminary (animal, cellular)
D — tradition & hypothesis

Ruler 2 — Action tier: ✅ do it / 🧪 worth trying / 📖 know it / 💸 don't spend money

Cross the two rulers and information becomes 3D: "a handful of nuts daily" is B/✅; "spermidine supplements" are C/💸 (food sources worth trying, supplements — save your money).

If a claim can't survive being graded, it doesn't get in.

Three entries you won't find in other books

1. Your brain cuts its own DNA to form memories (Cell 2015 → Nature 2023/2024)

To write a new memory, a neuron deliberately breaks DNA double strands near specific genes so that immediate-early genes can fire — then a repair complex (NPAS4–NuA4) stitches the cuts back. Neurons that go through this break-repair cycle get recruited into the memory circuit.

That's why learning a new framework feels like an overloaded server: it literally is one, physically rewiring. Give it compile time (sleep).

Grade: C (top-journal chain, animal mechanism) / 📖 know it.

2. Massaging your face and neck may help the brain drain waste (Nature 2025, mice)

The brain's metabolic waste rides cerebrospinal fluid out through a newly mapped route: orbital and nasal lymphatics converging into superficial cervical lymph vessels — right under the skin of your neck. In aged mice, gently pressing this channel with a calibrated device doubled CSF outflow.

The guide's massage routine follows this exact map. Why only grade C: the evidence is mice + a device. Nobody has tested human hands yet.

Grade: C / 🧪 worth trying (free, harmless, mechanism intact).

3. "Yang-qi ≈ mitochondria"? A food list the tradition never had (mapping D, foods B–C)

TCM describes yang-qi as warming, propelling, and protecting; mitochondria do thermogenesis, energy conversion, and quality control. Three-for-three (a Chinese team formally proposed this mapping in 2022 — hypothesis grade). Following the mapping, the guide prescribes a warming-food list the classics never had:

Pomegranate & walnuts → urolithin A, activates mitophagy (positive human RCTs as supplements; food conversion depends on gut flora)
Beets & spinach → dietary nitrate, makes human energy production more fuel-efficient (Larsen 2011)
Mushrooms → ergothioneine; higher blood levels associate with lower dementia risk in cohorts

The mapping is graded D. Every food on the list is graded separately on its own human evidence (B–C). The guide also offers a falsifiable prediction: if the mapping holds, people who self-report "yang deficiency" should show lower brown-fat activity — testable by PET-CT.

What's inside

10 chapters, structured on the 12 hallmarks of aging (Cell, 2023):

The Brain's Hidden Ops — lymphatic drainage, DNA breaks & memory, attention modulating inflammation, 40 Hz
Yang-qi & Mitochondria — the mapping, a new warming-food list, Zone 2, ergothioneine
Age Is Not a Number — facial biological age, methylation clocks, emotion & the genome
Old Formulas, New Ruler — a TCM formula, Baduanjin, vinegar, NAC, singing bowls — all scored on the same evidence scale
Ten Big Rocks of Longevity — the ten things that move lifespan most
A Long Day — morning light, standing up, screen use, the caffeine red line
Repair & Maintenance — sleep debt, screen curfew, workstation ergonomics
Focus Engineering — flow, working memory, pomodoro, meditation
Table & Supplements — nuts, omega-3, APOE4, spermidine (with "don't spend" flags)
The Twelve Pillars — the hallmarks map, the book's closing atlas (one pillar honestly left blank: "no lifestyle-level lever yet")

How to read it

GitHub (bilingual, English default): DeepoMe/Awesome-Programmer-Longevity
Mobile ebook (tap left/right to flip pages, adjustable font, dark mode): deepkang.com/ebook/ebook.html

License: CC BY-NC-ND. Issues welcome — challenge a grade, request a topic, point out what's over- or under-rated. The grading is an index that sends readers to the original papers, not a final verdict.

This guide is health education, not medical advice, diagnosis, or treatment. If you have a medical condition, take medication, or are pregnant/nursing, consult a physician before adopting any intervention. Individual variation is real; evidence grades describe population averages, not promises to you.

Update — this guide now runs a daily deep-dive series. One claim at a time, unpacked with its evidence grade:

#1: Your Brain Breaks Its Own DNA to Learn — Literally — learning physically cuts your genome; the survivors become the memory. (Grade C, Cell/Nature chain.)

#2: TCM Says “Yang Qi” Warms, Propels, and Protects. Your Mitochondria Do Exactly Those Three Things. — a 2,000-year-old TCM concept maps onto mitochondria with unsettling precision; the D-graded hypothesis that predicts a new warming-food list.
#3: Sitting Is Its Own Risk Factor, and Your Gym Hour Does Not Cancel It — sitting costs you on two separate accounts, and an hour at the gym does not pay off three unbroken hours in the chair. (Grade B.)
#4: The 2 P.M. Caffeine Rule: Your Afternoon Coffee Is Still in Your Blood at Midnight — a 5-hour half-life means a quarter of your 2 p.m. cup is still on your adenosine receptors at bedtime. You will not notice the hour it costs you. (Grade B.)

Follow @jxiong to catch the series.

Beyond Virtual Cells: SteeraMed Bench Gives AI the Map It Needs for Virtual Patients

JXIONG — Mon, 17 Aug 2026 06:48:04 +0000

SteeraMed Bench evaluates and optimizes human-scale biological maps for virtual patients—making representation selection a measurable driver of drug-repurposing decision quality.

tags: steeramed, ai4science, virtualpatients, drugdiscovery

Beyond Virtual Cells: SteeraMed Bench Gives AI the Map It Needs for Virtual Patients

AI can now predict protein structures, simulate cellular perturbations, and design molecules de novo. But ask it the question that actually determines whether a drug succeeds or fails—"How will this person's biology change under this intervention?"—and today's models go silent.

Molecules can be modeled. Cells can be simulated. But at the level of the person—precisely where drug-development value is won or lost—the infrastructure for systematic modeling remains absent.

This gap is what DeepoMe and The Future Laboratory at Tsinghua University are addressing with SteeraMed Bench, an evaluation and optimization engine for human-scale biological maps, released alongside a joint research paper on Preprints.org (ID 202608.0998, DOI: 10.20944/preprints202608.0998.v1).

The $900 Million Question: Why Phase II Is Where Value Lives

A recent Nature Reviews Drug Discovery Perspective lays out the economics: reducing failure by 20% at Phase II—the first test of whether a medicine delivers meaningful benefit in people—can save nearly $900 million per successful drug launch. The same improvement at early candidate selection has a comparatively small effect.

The article also finds that clinical programs using biomarkers to identify likely responders can cut the average cost per successful launch to roughly half that of unstratified programs.

In other words, the big money isn't lost in early "pick the molecule" decisions—it's lost at the human response stage. DeepoMe's positioning is clear: "We optimize the Phase II success-or-failure juncture—each successfully launched drug corresponds to nearly $900 million in value."

SteeraMed Bench: Context and Positioning

SteeraMed—a Steerable World Model framework for biomedicine—was first introduced by DeepoMe Chief Scientist Dr. Jianghui Xiong at the Beijing Zhiyuan Conference (BAAI). SteeraMed Bench is a new module within that framework: an evaluation-and-optimization engine that answers the question, "Which set of biological coordinates is most useful for the decision at hand?"

The Core Logic in Three Sentences

DeepoMe is building Human Response Intelligence, and its core logic can be stated in three steps: Use a steerable world model to compute virtual patients—predict whether a person will respond to an intervention. After the intervention, bring them back for retesting—whether the prediction was right is immediately verifiable. And this maps directly onto the most expensive, most valuable cut in drug development—Phase II clinical trials (each successfully launched drug corresponds to nearly $900 million; responder stratification can cut the cost per successful launch to approximately half).

Defining State: The Human-Scale Biological Map

Before you can predict how someone will respond, you need to define which biological dimensions matter for that person and that decision. DeepoMe organizes these into a human-scale biological map with two levels:

Modules: Reusable, evidence-linked units of the map—aging processes, immune state, organ function, nutrition, exposure history, and more.
Representations: Decision-ready combinations of modules selected for a particular disease, intervention, and response question.

A world model without an explicit map is like a navigator without a road network—no matter how good the rendering, it can't compute "from this state, under this intervention, where do I arrive?"

Reasoning over explicit, testable representations of human state, mechanism, and intervention is what makes a Biomedical World Model genuinely steerable.

SteeraMed Bench: Methodology and Key Results

Once you have a map, the methodological question follows: among all modules and their combinations, which map is most useful for the current decision?

SteeraMed Bench provides a measurable answer: it tests which modules and combinations improve intervention reasoning, then selects and integrates the strongest-performing combinations into optimized maps.

In the reported analyses:

Parameter	Value
Total modules	332 (aging hallmarks, organ systems, immunity, nutrition, food-derived interventions, TCM syndrome proxies, food-as-medicine targets)
Small molecules	1,916 DrugBank compounds
Disease tasks	5 chronic diseases, extrapolated to 23 disease categories
Core metric	recall\@20 (proportion of known drug-disease pairs captured in top-20 ranked candidates)

Three core results emerged:

1. No single biological map is universally optimal. The full 332-module atlas is strictly optimal in only 9 of 23 disease categories. Different diseases have different best-fitting coordinate combinations: extended aging hallmarks lead in type 2 diabetes and osteoporosis; food-as-medicine combinations are strongest for depression; nutrient modules dominate atherosclerosis/hyperlipidemia.

2. Representation selection is a measurable driver of decision quality. Just 117 nutrient and extended modules achieve an average recall\@20 of 0.494 across five diseases—already approaching the full 332-module atlas at 0.524. This fact was previously judged mostly by expert intuition and default single-atlas assumptions.

3. Negative results are reported honestly. LLM-proposed candidate modules showed no significant improvement after multiple-testing correction; leave-one-disease-out validation performed near random. This is methodologically important: it demonstrates the benchmark does not stack the deck for any particular approach.

Quantitative Stress-Testing of Knowledge Frameworks

The framework also includes an LLM-assisted workflow for proposing and evaluating candidate modules against explicit utility and redundancy criteria.

Consider the Hallmarks of Aging framework: it has expanded from 9 to 12 and now to 14 major hallmarks. Such scientific frameworks traditionally evolve through expert synthesis of growing literature. SteeraMed Bench adds a quantitative layer—it tests whether a proposed hallmark is a genuinely independent dimension that improves decision quality, or merely redundant with existing representations.

For researchers, this is a new instrument: a ruler for scoring knowledge frameworks.

# Conceptual: How SteeraMed Bench evaluates a candidate module
def evaluate_module(candidate_module, existing_modules, task):
    """
    Test whether adding candidate_module improves decision quality
    beyond what existing_modules already provide.
    """
    baseline_recall = compute_recall(existing_modules, task)
    augmented_recall = compute_recall(existing_modules + [candidate_module], task)
    delta = augmented_recall - baseline_recall
    is_significant = multiple_testing_correction(delta, alpha=0.05)
    return {
        "module": candidate_module,
        "delta_recall": delta,
        "is_significant": is_significant,
        "verdict": "ADD" if is_significant else "REDUNDANT"
    }

Expert Perspective

"AI for Science has become remarkably good at seeing molecules and cells. But before we can ask how an intervention changes a person, we must define the biological state that matters for that person and that decision. Human Response Intelligence has two tasks: first, identify the dimensions that best represent human state; second, learn how that state may change under intervention. SteeraMed Bench makes the first task measurable by evaluating and optimizing modules into the maps that provide the strongest basis for intervention reasoning."
— Dr. Jianghui Xiong, Chief Scientist, DeepoMe

Why Longevity Science as the First Domain

For DeepoMe, longevity science and aging-related disease are the initial learning domain for this AI mission. These settings combine four conditions that make human-biology learning feasible:

Repeated measurement: Biological state can be sampled multiple times
Multidimensional state: Multiple biological axes can be characterized simultaneously
Modifiable interventions: Diet, supplements, drugs, lifestyle can be documented
Cross-disease relevance: Aging processes span multiple disease categories

DeepoMe's multidimensional aging-assessment work, including Capome, is an entry point for measuring human biological state longitudinally and interpretably—linking repeated measurements with structured intervention exposure and follow-up creates the conditions to learn how, and for whom, an intervention changes that state.

Toward N-of-1 Prediction

For research and biopharma partners, DeepoMe is developing an auditable translational-decision workflow for:

Prioritizing drug-repurposing hypotheses
Defining candidate responder populations
Specifying the next validation experiment or longitudinal study

For aging and Longevity Medicine, the long-term significance is moving intervention reasoning from population-level associations toward individualized N-of-1 prediction and, ultimately, the generation of individualized intervention programs.

A "virtual patient" is not a rendered human image, nor a declared-complete digital twin—it is a testable, auditable inference substrate: identify the biological dimensions most relevant to a given individual, prioritize intervention hypotheses, pair them with longitudinal follow-up, and verify whether predictions hold.

The next phase of world models may not be about parameter scale—it may be about who first accumulates auditable, testable human-representation assets.

References

Xiong J, Xia Q. Toward a Self-Learning AI Agent for Drug Repurposing: Building Human-Scale Representations for Virtual Patients. Preprints, 2026. DOI: 10.20944/preprints202608.0998.v1. [This paper]
López-Otín C, Blasco MA, Partridge L, et al. The hallmarks of aging. Cell, 2013, 153(6): 1194-1217.
López-Otín C, Blasco MA, Partridge L, et al. Hallmarks of aging: An expanding universe. Cell, 2024, 187(2): 361-378.

Disclaimer: This article is a technology overview. The cited research has not yet completed peer review and does not constitute medical advice. For health concerns, consult a qualified healthcare professional.

Building a Biomedical World Model in Python: SteeraMed Core

JXIONG — Mon, 01 Jun 2026 05:44:14 +0000

What if you could ask: "Which compound is most likely to reverse this specific patient's molecular aging?" — and get a 4-layer auditable evidence chain, not a black-box recommendation?

That's what SteeraMed Core does. It's an open-source Python package that applies the "world model" concept from reinforcement learning to biomedicine: quantify how an individual's biology deviates from normal, then simulate which compounds can steer it back.

No PyTorch. No TensorFlow. No GPU. Just numpy, pandas, scipy, and matplotlib.

pip install steeramed-core

Why This Matters

Epigenetic clocks (Horvath 2013, Hannum 2013) can measure your "biological age" from DNA methylation. But measuring aging is only step one — the real challenge is intervention: how do you steer molecular state back toward a younger profile?

The Hallmarks of Aging framework (López-Otín et al., Cell 2013, >15k citations) defined 9 hallmarks of aging, expanded to 12 in the 2023 update (Hallmarks of aging: An expanding universe). SteeraMed operationalizes this framework using the MSigDB Hallmark 50 gene sets (Liberzon et al., Cell Systems 2015) — 50 curated biological pathway gene sets covering aging, cancer, immunity, and metabolism.

Note: MSigDB "Hallmark" (50 pathway gene sets) ≠ Hallmarks of Aging (12 aging pillars). SteeraMed uses the former as its functional module definitions.

What Is a Biomedical World Model?

In reinforcement learning, a world model (Ha & Schmidhuber, 2018) is an internal simulator that predicts the consequences of different actions. Think of AlphaGo mentally simulating "if I play here, what will my opponent do?"

Applied to biomedicine:

State representation — Quantify how an individual deviates across 50 biological pathway modules using DNA methylation
Action simulation — Simulate "if we apply compound X, which disrupted modules get corrected" on the PPI network
Auditable reasoning — Generate a 4-layer traceable evidence chain, not a black-box output

	Traditional Systems Biology / AI Drug Discovery	Biomedical World Model
Unit of analysis	Population mean	Individual (N-of-1)
Inference direction	Forward (drug → effect)	Reverse (deviation → corrective drug)
Output	Drug repurposing candidates	4-layer individualized evidence chain
Confidence	Clinical trial statistics	Bootstrap resampling confidence

Project Architecture

steeramed_core/
├── __init__.py              # Entry: EvidenceChain + load_example_patient
├── __main__.py              # CLI: interactive case selector + batch mode
├── core/                    # Core algorithms
│   ├── config.py            # Global config + disease presets
│   ├── delta.py             # N-of-1 delta vector computation
│   ├── evidence_chain.py    # 4-layer evidence chain data structures
│   └── semo.py              # SA scoring + compound ranking
├── presets/                 # Pre-computed data (3 real clinical cases)
│   ├── catalog.json         # Case catalog
│   ├── datasets.json        # GEO dataset metadata
│   ├── positive_controls.json # Known drug ground truth
│   └── example_patients/    # 3 JSON patient files
├── viz/                     # Visualization (Nature-style theme)
│   ├── theme.py             # Color palette + rcParams
│   ├── hallmark_bar.py      # Hallmark perturbation bar chart
│   ├── drug_ranking.py      # Top-10 compound ranking chart
│   ├── evidence_network.py  # Drug-PPI-Hallmark network graph
│   ├── patient_card.py      # Single-page patient summary card
│   ├── evidence_view.py     # Scientist view (Paper Fig 6/8)
│   └── patient_view.py      # Patient view (Paper Fig 4/7)
└── examples/                # Reproduction scripts
    ├── reproduce_aging_patient_view.py
    ├── reproduce_ra_evidence_chain.py
    └── reproduce_dep_evidence_chain.py

Minimal dependencies:

dependencies = [
    "numpy>=1.21",
    "pandas>=1.3",
    "scipy>=1.7",
    "matplotlib>=3.5",
]

The 4-Layer Evidence Chain

Layer 1: Which Hallmark Pathways Are Disrupted?

Map DNA methylation to PPI (protein-protein interaction) network modules. Evaluate all 50 modules, find the ones significantly deviating from age-matched controls.

# core/delta.py — N-of-1 Delta vector
def compute_n1_delta(patient_genes, control_genes):
    """
    Δ_i = x_i - x̄_matched
    Matched controls: same sex, age ±5 years, K=10
    """
    matched_mean = control_genes.mean(axis=0)
    return patient_genes - matched_mean

Example findings in the aging case:

NAD+ metabolism module disrupted (Loss of NAD+)
Inflammatory modules upregulated (TNFα/NF-κB, IL-6/JAK/STAT3)
Protein homeostasis disturbed (Unfolded Protein Response)
Some modules remain normal (Hedgehog, Notch signaling)

Layer 2: Which Compounds Can "Steer Back"?

Compute a Steerability Alignment Score (SA Score) — essentially a Welch-type contrast statistic comparing methylation deltas of compound target genes vs. non-target genes.

# core/semo.py
def compute_sa_score(delta, target_genes, all_genes):
    """
    SA Score = Welch t-statistic
    Compare compound target genes vs non-targets in disrupted modules
    """
    target_delta = delta[delta.index.isin(target_genes)]
    non_target_delta = delta[~delta.index.isin(target_genes)]
    return welch_t(target_delta, non_target_delta)

Compound-target data comes from STITCH database. The ranking uses importance voting across bootstrap samples:

def rank_compounds_by_importance(sa_matrix, compounds):
    """
    Importance Voting: each sample's top-1 compound gets 1 vote
    """
    votes = defaultdict(int)
    for sample_sa in sa_matrix:
        top_compound = sample_sa.idxmax()
        votes[top_compound] += 1
    return sorted(votes.items(), key=lambda x: -x[1])

Aging case results: Niacin #1 (targets NAD+ metabolism), Colchicine #2 (anti-inflammatory). 2/5 top hits are known geroprotectors.

Layer 3: Mechanism Traceability

Trace each compound's mechanism: compound targets → PPI network neighbors → hub genes → corresponding Hallmark pathway.

Example: Niacin → NAMPT/NAPRT → NAD+ metabolism module → Loss of NAD+

Layer 4: Bootstrap Confidence

1000 bootstrap resamples to test ranking stability. Top-1 compound retention rate determines the evidence level:

Level	Bootstrap Stability	Meaning
STRONG	≥80%	Robust recommendation
MODERATE	50-80%	Reasonable evidence
EXPLORATORY	<50%	Hypothesis-generating only

The evidence chain is a clean dataclass:

# core/evidence_chain.py
@dataclass
class EvidenceChain:
    patient_id: str
    disease: str
    perturbed_modules: List[PPIModule]     # Layer 1
    top_compounds: List[CompoundMatch]     # Layer 2
    mechanism_map: dict                    # Layer 3
    bootstrap_stability: dict              # Layer 4

    def summary(self) -> str: ...
    def to_dict(self) -> dict: ...
    def to_json(self, path: str) -> None: ...

    @classmethod
    def from_dict(cls, data: dict) -> 'EvidenceChain':
        # Backward compatible: ignores unknown fields
        ...

Quick Start

Current version: SteeraMed Core is a proof-of-concept demo with 3 built-in real clinical cases from GEO. Custom data upload (450K/EPIC methylation arrays) is coming in future releases. Follow updates at steerable.world.

Interactive Mode

pip install steeramed-core
python -m steeramed_core

SteeraMed Core — N-of-1 Evidence Chain Explorer
=================================================

Select a patient case:
  [1] Aging · Population Screening
  [2] RA · 51M · T-cell Perturbation
  [3] Depression · 52M · Innate Immunity

Enter choice [1-3]:

Batch Mode

python -m steeramed_core --all             # Generate all cases
python -m steeramed_core --case ra_303     # Specific case
python -m steeramed_core --list            # List available cases

Python API — Load & Inspect

from steeramed_core import EvidenceChain, load_example_patient

patient = load_example_patient("ra_patient_303")
print(patient.summary())

# Inspect the 4 layers
print(f"Disrupted modules: {len(patient.perturbed_modules)}")
print(f"Top compound: {patient.top_compounds[0].compound_id}")
print(f"Bootstrap stability: {patient.bootstrap_stability}")

Python API — Generate All Charts

from steeramed_core import load_example_patient
from steeramed_core.viz.patient_card import plot_patient_card
from steeramed_core.viz.drug_ranking import plot_drug_ranking
from steeramed_core.viz.hallmark_bar import plot_hallmark_bar
from steeramed_core.viz.evidence_network import plot_evidence_network

data = load_example_patient("ra_patient_303").to_dict()

# 4 charts: hallmark perturbation, drug ranking, network, patient card
for fn, name in [
    (plot_hallmark_bar, "hallmark_bar"),
    (plot_drug_ranking, "drug_ranking"),
    (plot_evidence_network, "evidence_network"),
    (plot_patient_card, "patient_card"),
]:
    fig = fn(data)
    fig.savefig(f"{name}.png", dpi=300, bbox_inches="tight")

Python API — Publication-Grade Figures

from steeramed_core.viz.evidence_view import plot_evidence_chain
from steeramed_core.viz.patient_view import plot_patient_view

# Scientist view — 3-panel evidence chain (Paper Fig 6/8 style)
fig = plot_evidence_chain(data)
fig.savefig("evidence_chain.png", dpi=300, bbox_inches="tight")

# Patient view — 3-panel card (Paper Fig 4/7 style)
fig = plot_patient_view(data)
fig.savefig("patient_view.png", dpi=300, bbox_inches="tight")

Validation Results

Retrospective positive control validation on 3 GEO datasets:

Cohort	Disease	N	Key Finding	Evidence
GSE40279 (Hannum)	Aging	656	Niacin #1, 2/5 geroprotectors	MODERATE
GSE42861	Rheumatoid Arthritis	689	6/10 known RA drugs recovered, pentoxifylline #1	STRONG
GSE128235	Depression (MDD)	533	Creatine #1, innate immunity dominant	EXPLORATORY

The RA cohort is the strongest result: known RA drugs recovered at 5.8x random expectation, confirming that PPI module-level alignment captures meaningful drug-disease matches.

The depression cohort's top-1 compound (creatine) has only 24.5% bootstrap stability — honestly flagged as EXPLORATORY. This reflects the high heterogeneity and weak methylation signal in MDD.

Configuration

All hyperparameters live in core/config.py:

# PPI network
PPI_SCORE_CUTOFF = 400    # STRING confidence threshold
PPI_MIN_SIZE = 20         # Min genes per module
PPI_MAX_SIZE = 800        # Max genes per module

# Compound targets
STITCH_SCORE = 200        # STITCH confidence threshold
# Target count range: [60, 300]

# Bootstrap
BOOTSTRAP_N1 = 200        # N-of-1 resampling iterations
BOOTSTRAP_GROUP = 100     # Group-level iterations

# Matching
MATCH_K = 10              # Number of matched controls
MATCH_CALIPER = 5         # Age matching window (years)

Visualization API

Function	Chart Type	Size	Purpose
`plot_hallmark_bar()`	Horizontal bar	8xN in	Hallmark perturbation magnitude
`plot_drug_ranking()`	Horizontal bar	8xN in	Top-10 compound ranking
`plot_evidence_network()`	Bipartite network	10x7 in	Drug-Hallmark alignment
`plot_patient_card()`	Card layout	8x10 in	Single-page patient summary
`plot_evidence_chain()`	3-panel	7.2x9.0 in	Scientist view (Paper Fig 6/8)
`plot_patient_view()`	3-panel card	7.5x10.5 in	Patient view (Paper Fig 4/7)

All viz functions use matplotlib.use('Agg') — works on headless servers and CI.

Honest Limitations

Retrospective validation only — positive control recovery, not prospective clinical trials
Bootstrap confidence varies — depression case is only 24.5% stable (EXPLORATORY)
Single omics — DNA methylation only; transcriptomics/proteomics coming in future versions
Simple matching — age ±5 years + sex, doesn't control for cell composition
Demo stage — 3 preset cases, custom data upload coming soon

References

López-Otín C, et al. The hallmarks of aging. Cell, 2013, 153(6): 1194-1217.
López-Otín C, et al. Hallmarks of aging: An expanding universe. Cell, 2023, 186(2): 243-278.
Horvath S. DNA methylation age of human tissues and cell types. Genome Biology, 2013, 14: R115.
Hannum G, et al. Genome-wide methylation profiles reveal quantitative views of human aging rates. Molecular Cell, 2013, 49(4): 621-635.
Ha D, Schmidhuber J. World models. NeurIPS, 2018.
Liberzon A, et al. The Molecular Signatures Database Hallmark Gene Set Collection. Cell Systems, 2015, 1: 417-425.
Xiong J. World Models for Biomedicine: A Steerability Framework. Preprints.org, 2026. DOI: 10.20944/preprints202605.0366.v1
Xiong J. SteeraMed: A Biomedical World Model for N-of-1 Intervention Reasoning. Preprints.org, 2026. DOI: 10.20944/preprints202605.1578.v1

Medical World Model Prototype: SteeraMed Explained — From Report Reading to Life Trajectory Reasoning

JXIONG — Sat, 30 May 2026 02:24:47 +0000

TL;DR

Over the past few years, many medical AI applications have focused on two major categories of tasks:

Recognition and prediction: medical imaging, assisted diagnosis, risk scoring, and report interpretation.
Generation and explanation: clinical note summarization, health Q&A, medical knowledge retrieval, and report explanation.

These tasks are useful, but chronic disease management, aging intervention, and longevity medicine require something more difficult:

How does an individual biological state change over time under the influence of lifestyle, drugs, nutrition, exercise, sleep, stress, and environmental exposure?

This is not just an NLP question-answering problem. It is not simply a single-point classification problem either.

It is closer to a dynamic system modeling problem.

That is why medical AI may need to move from report interpretation toward life trajectory reasoning.

In other words, we need a new framework: the medical world model.

This article discusses:

Why report-reading medical AI is not enough.
What a medical world model is.
How State-Action-Transition-Evidence-Feedback can describe biomedical reasoning.
How SteeraMed Core works as an early medical world model prototype.
How the SEMO algorithm supports candidate intervention reasoning.
Current limitations and engineering challenges for medical world models.

1. Why report-reading AI is not enough

Medical AI can already do many useful things:

Read health checkup reports.
Explain laboratory values.
Summarize clinical notes.
Answer health-related questions.
Retrieve medical knowledge from literature.
Support imaging interpretation and risk prediction.

These are valuable applications.

However, when we move into chronic disease management, aging intervention, and longevity medicine, the problem becomes more complex.

An individual may have many types of data at the same time:

clinical biomarkers;
DNA methylation data;
transcriptomics, proteomics, metabolomics, and other multi-omics data;
inflammation, immune, and metabolic markers;
sleep, exercise, stress, and lifestyle records;
symptoms and functional status;
medication, nutrition, exercise, and psychological intervention records;
retesting and follow-up feedback.

The hard question is not:

What does this biomarker mean?

The hard question is:

How are these signals connected?
Which changes are primary and which are secondary?
Which state deviations should be tracked over time?
Which actions may change the future state?
Which metrics should be used for feedback?
How should the next round of reasoning be updated after feedback?

This is not just an NLP task.

It is a state transition problem.

This is exactly the type of problem a medical world model is meant to address.

2. What is a medical world model?

In AI and reinforcement learning, a world model usually refers to an internal model of how the world changes.

A simplified expression is:

state_t + action_t -> state_t+1

In plain language:

current state + current action -> next state

If we translate this idea into medicine and life sciences, the problem becomes:

patient_state_t + intervention_action_t -> patient_state_t+1

But the human body is not a game environment.

Medical systems have several important differences:

The human body is multi-scale: molecules, cells, tissues, organs, and behaviors all interact.
Medical data is multi-modal: omics, clinical biomarkers, imaging, text, wearables, and follow-up records.
Health trajectories are long-term: chronic disease progression, aging, and intervention responses may unfold over years.
The state is partially observable: many important variables cannot be directly measured.
Interventions have safety constraints: model reasoning must not be treated as clinical decision-making.

Therefore, a medical world model should not be understood as a universal predictor.

A more useful definition is:

A medical world model is a structured reasoning framework that represents biomedical state, candidate action, transition hypothesis, evidence chain, and feedback loop over time.

A medical world model does not merely ask:

Is this person high risk?

It asks:

What is the current state?
What are the candidate actions?
How may an action change the state?
What evidence supports the transition hypothesis?
What feedback should be observed?
How should the next round of reasoning be updated?

3. The core abstraction: State-Action-Transition-Evidence-Feedback

To make a medical world model more engineering-friendly, we can start with five core objects:

State      -> current biological and clinical state
Action     -> candidate intervention or management action
Transition -> state-change hypothesis
Evidence   -> mechanism and literature evidence
Feedback   -> retesting and follow-up observations

A simplified Python-style representation might look like this:

class MedicalWorldModel:
    def __init__(self):
        self.state = None       # current patient state
        self.actions = []       # candidate actions
        self.transitions = []   # transition hypotheses
        self.evidence = []      # evidence chains
        self.feedback = []      # follow-up observations

A more detailed abstraction:

class State:
    biomarkers: dict
    omics_features: dict
    clinical_features: dict
    functional_features: dict
    lifestyle_features: dict

class Action:
    action_type: str
    target_modules: list
    mechanism_hypothesis: str
    constraints: list

class Transition:
    from_state: State
    action: Action
    expected_direction: dict
    uncertainty: float

class Evidence:
    source_type: str
    source_url: str
    mechanism: str
    confidence_level: str

class Feedback:
    followup_time: str
    observed_changes: dict
    adherence: str
    adverse_events: list

This is not a clinical implementation.

It is a conceptual data structure showing one important principle:

A medical world model should not only generate a text recommendation. It should organize state, action, transition, evidence, and feedback into a traceable structure.

This is also the foundation of a steerable medical world model.

4. Medical world model vs. risk prediction model

Traditional medical AI often focuses on risk prediction.

A risk model usually asks:

What is the probability that this person will develop a disease?

or:

Is this person high-risk, medium-risk, or low-risk?

A medical world model asks a different set of questions:

If we take a certain action, how may the state change?
Why may the state change in that direction?
Which biomarkers or functional outcomes should be observed?
How should feedback update the next round of reasoning?

A simple comparison:

Dimension	Risk prediction model	Medical world model
Core question	How high is the risk?	How may the state change?
Input	Current features	State, action, evidence, feedback
Output	Risk score or class	Transition hypothesis and observation plan
Time dimension	Often weak	Central
Feedback loop	Usually secondary	Core component

For chronic disease management and aging intervention, the second approach is especially important.

Aging is not a single disease. Chronic disease is rarely a single pathway. Both are multi-system, long-term, and feedback-sensitive processes.

5. From a steerable world model framework to SteeraMed Core

Based on this idea, DeepOMe first proposed a steerable world model framework for biomedicine.

In the preprint World Models for Biomedicine: A Steerability Framework, we explored the following question:

Can biological state only be passively predicted, or can it be modeled around observable state, executable action, transition hypothesis, evidence chain, and feedback process in a steerable way?

Preprint DOI / URL:

https://doi.org/10.20944/preprints202605.0366.v1

Based on this framework, DeepOMe released another preprint on May 25, 2026:

SteeraMed: A Biomedical World Model for N-of-1 Intervention Reasoning Across Chronic Diseases and Aging

This preprint further implements the steerable biomedical world model idea into an early tool form for chronic disease and aging intervention scenarios: SteeraMed Core.

Preprint DOI / URL:

https://doi.org/10.20944/preprints202605.1578.v1

It is important to be clear about the boundary:

SteeraMed Core is not a clinical decision system. It is an early medical world model prototype for research, clinical collaboration, and health management tool exploration.

6. SteeraMed Core as a medical world model prototype

From an engineering point of view, SteeraMed Core can be understood as an early medical world model prototype.

Its goal is not to directly output a treatment plan.

Its goal is to structure the reasoning process.

A simplified pipeline:

Input:
    Individual multi-dimensional data
        - DNA methylation
        - multi-omics features
        - clinical biomarkers
        - symptoms and functional status
        - lifestyle records

Process:
    1. Build state representation
    2. Detect state deviation
    3. Generate candidate actions
    4. Build transition hypotheses
    5. Link evidence chains
    6. Design feedback metrics

Output:
    Explainable, traceable, auditable intervention reasoning structure

A simplified pseudocode version:

def steeramed_core_pipeline(user_data):
    state = build_state_representation(user_data)
    deviations = detect_state_deviation(state)
    candidate_actions = generate_candidate_actions(deviations)

    transition_hypotheses = []

    for action in candidate_actions:
        evidence_chain = retrieve_evidence(state, action)
        transition = infer_transition(state, action, evidence_chain)
        feedback_plan = design_feedback_metrics(state, action, transition)

        transition_hypotheses.append({
            "state": state,
            "action": action,
            "evidence": evidence_chain,
            "transition": transition,
            "feedback": feedback_plan,
        })

    return transition_hypotheses

A real system would be much more complex.

It would involve:

data cleaning;
feature engineering;
network modeling;
evidence retrieval;
uncertainty estimation;
human-in-the-loop review;
longitudinal feedback integration.

But the core idea is simple:

Do not just generate recommendations. Build a traceable reasoning structure.

That is what makes SteeraMed Core a medical world model prototype rather than a report generator.

7. SEMO: the network-remodeling layer behind SteeraMed

One important algorithmic foundation behind SteeraMed Core is the SEMO algorithm.

SEMO stands for:

Selective Remodeling of Protein Networks by Chemicals

It is a computational framework for selectively remodeling protein networks using chemicals.

The key question SEMO tries to address is:

In an individual biological network, where are the measurable state deviations? Which candidate compounds, nutrients, drugs, or lifestyle actions may push the state in a better direction through network remodeling?

From an algorithmic perspective, SEMO is not a simple recommender system.

A simple recommender system might look like this:

user features -> recommended supplement or lifestyle advice

SEMO emphasizes a different chain:

individual omics state
-> network deviation modules
-> candidate action networks
-> mechanism evidence
-> feedback metrics

In other words, SEMO focuses on:

network deviation in individual omics signals;
target networks associated with candidate actions;
mechanism connections between state and intervention;
feedback metrics that can track the intervention hypothesis.

The related method framework has been granted an invention patent.

The relationship between SEMO and SteeraMed Core can be summarized as:

SEMO provides the network-remodeling and candidate-intervention reasoning layer. SteeraMed Core places this capability inside a medical world model framework of State-Action-Transition-Evidence-Feedback.

8. Why medical world models matter for longevity AI

Longevity technology should not stop at “measuring an age.”

If a test cannot explain state deviation, organize candidate actions, design feedback metrics, and support the next round of calibration, it may easily become a one-time report product.

A meaningful AI longevity system may need at least four layers:

1. Data input layer
   DNA methylation, multi-omics, clinical biomarkers,
   wearable data, symptoms, lifestyle, follow-up records

2. State representation layer
   organ systems, molecular pathways, functional states,
   risk signals, aging hallmarks

3. Action reasoning layer
   nutrition, exercise, drugs, psychology, sleep,
   lifestyle and other candidate actions

4. Feedback calibration layer
   retesting, follow-up, adherence, adverse events,
   state update and next-round reasoning

From this perspective, a medical world model is not just an app.

It is not just a report generator.

It is a long-term infrastructure for organizing state, action, transition, evidence, and feedback.

9. Three early demo views

SteeraMed Core currently includes three early demo views to show how a medical world model may move from concept to prototype.

The descriptions below are intentionally brief. For detailed case definitions, data representations, evidence-chain construction, and prototype outputs, please refer to the SteeraMed preprint:

SteeraMed: A Biomedical World Model for N-of-1 Intervention Reasoning Across Chronic Diseases and Aging

DOI: 10.20944/preprints202605.1578.v1

9.1 Depression Patient View

Depression may involve sleep, stress, inflammation, metabolism, neuroendocrine state, life events, drug response, social support, and individual biological differences.

In the SteeraMed preprint, the Depression Patient View is used as one example of how a medical world model can organize patient-facing state signals, candidate actions, evidence chains, and follow-up feedback into a computable observation path. The technical details are described in the SteeraMed preprint above.

9.2 Anti-aging Patient View

In longevity medicine, individual differences are especially important.

In the SteeraMed preprint, the Anti-aging Patient View demonstrates how DNA methylation, Aging Hallmarks, organ-system signals, metabolic-immune status, and lifestyle factors can be integrated into a medical world model view to help identify candidate steerability factors and retesting metrics. The detailed implementation is described in the SteeraMed preprint above.

9.3 Rheumatoid Arthritis Scientist View

Rheumatoid arthritis is a chronic inflammatory and immune-related disease involving immune cells, inflammatory cytokines, joint tissue damage, drug response, and long-term disease activity.

In the SteeraMed preprint, the Rheumatoid Arthritis Scientist View is used to illustrate how a medical world model can organize disease state, immune-inflammatory modules, candidate targets, evidence chains, and feedback metrics into an auditable research structure. The detailed case description is provided in the SteeraMed preprint above.

10. Current limitations and engineering challenges

Medical world models are promising, but they must be treated carefully.

Several challenges remain:

Data challenge: real-world data is often discontinuous, incomplete, and non-standardized.
Validation challenge: model reasoning requires prospective observation and long-term follow-up.
Causality challenge: correlation is not causation, and mechanism chains still need validation.
Safety challenge: model outputs must not be treated as direct clinical decisions.
Engineering challenge: multi-omics, clinical, text, and wearable data need unified modeling.
Human-in-the-loop challenge: the system should support researchers and professionals, not replace professional judgment.

Therefore, SteeraMed Core is still an MVP tool.

It is mainly intended for research, clinical collaboration, and health management tool exploration. It is used to organize evidence, generate hypotheses, and track feedback.

It does not replace physicians, does not promise individual clinical outcomes, and should not be considered a validated medical device, diagnostic system, or treatment decision system.

11. Takeaway

If medical AI wants to truly enter chronic disease management and longevity medicine, it cannot remain limited to static report interpretation.

It needs to address a harder set of questions:

How does state change?
How does action influence state?
How does evidence constrain transition reasoning?
How does feedback calibrate the model?

That is the value of a medical world model.

SteeraMed Core is an early medical world model prototype developed by DeepOMe. It attempts to organize State-Action-Transition-Evidence-Feedback into an explainable, traceable, auditable, and iterative reasoning framework for chronic disease management, aging intervention, and N-of-1 reasoning.

About DeepOMe

DeepOMe starts from DNA methylation and aims to build AI world models of human aging, moving longevity technology from static testing toward dynamic reasoning.

DeepOMe uses DNA methylation and other epigenetic information as an entry point for representing life state, exploring molecular aging states, organ-system risk signals, and long-term health trajectory modeling.

In AI longevity technology, DeepOMe focuses on SEMO, SteeraMed Core, the DeepKang longevity medicine agent platform, and root-cause reasoning agents for individualized health management, resource integration, and long-term feedback loops.

References

[1] Ha, D., & Schmidhuber, J. World Models. arXiv, 2018.

https://arxiv.org/abs/1803.10122

Conference version: Recurrent World Models Facilitate Policy Evolution. NeurIPS, 2018.

https://papers.nips.cc/paper/7512-recurrent-world-models-facilitate-policy-evolution

[2] LeCun, Y. A Path Towards Autonomous Machine Intelligence. OpenReview, 2022.

https://openreview.net/forum?id=BZ5a1r-kVsf

[3] Moor, M., Banerjee, O., Abad, Z. S. H., Krumholz, H. M., Leskovec, J., Topol, E. J., & Rajpurkar, P. Foundation models for generalist medical artificial intelligence. Nature, 2023, 616, 259–265.

https://doi.org/10.1038/s41586-023-05881-4

[4] Katsoulakis, E., Wang, Q., Wu, H., Shahriyari, L., Fletcher, R., Liu, J., Achenie, L., Liu, H. H., Jackson, P., Xiao, Y., Syeda-Mahmood, T., & Deng, J. Digital twins for health: a scoping review. npj Digital Medicine, 2024, 7, 77.

https://doi.org/10.1038/s41746-024-01073-0

[5] Xiong, J. World Models for Biomedicine: A Steerability Framework. Preprints.org, 2026.

https://doi.org/10.20944/preprints202605.0366.v1

[6] Xiong, J. SteeraMed: A Biomedical World Model for N-of-1 Intervention Reasoning Across Chronic Diseases and Aging. Preprints.org, 2026.

DOI: 10.20944/preprints202605.1578.v1

[7] Xiong, J. Utilizing Pre-trained Network Medicine Models for Generating Biomarkers, Targets, Re-purposing Drugs, and Personalized Therapeutic Regimes: COVID-19 Applications. bioRxiv, 2023.

https://doi.org/10.1101/2023.02.21.527754

Why Systems Biology Is Not the Same as a Medical World Model

JXIONG — Thu, 21 May 2026 13:56:41 +0000

If you have worked with computational biology, multi-omics analysis, pathway modeling, biomedical knowledge graphs, or systems biology, the phrase medical world model may sound suspicious at first.

You might ask:

Isn't this just systems biology with a new name?

That is a fair question.

Life sciences have been modeling gene regulatory networks, signaling pathways, metabolic systems, disease mechanisms, and biological perturbations for decades. If a medical world model were only a rebranding of systems biology, digital twins, or large language models, it would not add much technical value.

But from a system-design perspective, the difference is not mainly about terminology.

It is about object boundaries.

A systems biology model often focuses on:

component -> relation -> pathway -> network -> mechanism

A medical world model needs to additionally represent:

state + action + evidence -> transition hypothesis -> feedback update

In other words:

systems biology helps us understand how biological systems work;
medical world models aim to reason about how an individual state may change under a defined action;
steerable medical world models further add objectives, constraints, safety gates, human review, feedback, and auditability.

This article explains the distinction from a developer's perspective.

1. Three model types, three engineering boundaries

A useful way to avoid confusion is to compare prediction models, systems biology models, and medical world models.

Model type	Core question	Main objects	Typical output	Engineering keywords
Prediction model	How high is the future risk?	features, labels	risk score, class, probability	classification / regression
Systems biology model	How does the biological system work?	genes, proteins, pathways, networks	mechanism, network dynamics	graph / ODE / network model
Medical world model	How may state change after an action?	state, action, transition, evidence, feedback	transition hypothesis, audit trail	state-action-feedback loop

A standard medical risk model may look like:

risk = predict_risk(patient_features)

A systems biology model may look like:

network_state = simulate_pathway_dynamics(
    pathway_graph,
    initial_conditions,
    perturbation
)

A medical world model is closer to:

transition = estimate_transition_hypothesis(
    state=current_patient_state,
    action=candidate_intervention,
    evidence=evidence_chain
)

feedback = collect_feedback(
    patient_id=patient_id,
    action=candidate_intervention,
    time_window_weeks=8
)

updated_state = update_state(
    previous_state=current_patient_state,
    action=candidate_intervention,
    transition=transition,
    feedback=feedback
)

The important difference is not that one model is more complex than another.

The difference is the modeling target:

feature prediction
    -> mechanism modeling
    -> action-conditioned transition reasoning

2. Systems biology can model perturbations, but action is not always a decision object

Systems biology is not action-free.

It can model:

gene knockout;
drug perturbation;
pathway activation or inhibition;
environmental change;
ODE-based dynamics;
network control;
multi-omics perturbation response.

So this statement would be wrong:

systems biology has no action
medical world models have action

A better distinction is:

Systems biology can model perturbations and responses. A medical world model needs to turn action into a structured decision object and place it inside an evidence, transition, feedback, safety, and audit loop.

In systems biology, a perturbation may be an input parameter:

result = simulate_network(
    graph=pathway_graph,
    perturbation={"gene_x": "knockout"}
)

In a medical world model, an action is not just a perturbation parameter. It must be executable, recordable, auditable, monitorable, and feedback-compatible.

For example:

from dataclasses import dataclass
from typing import List

@dataclass
class InterventionAction:
    action_id: str
    category: str
    description: str
    target_mechanisms: List[str]
    intensity: str
    duration_weeks: int
    monitoring_markers: List[str]
    safety_constraints: List[str]

Example:

action = InterventionAction(
    action_id="nutrition_low_glycemic_8w",
    category="nutrition",
    description="8-week low-glycemic dietary adjustment",
    target_mechanisms=[
        "postprandial_glucose_variability",
        "insulin_resistance",
        "weight_management"
    ],
    intensity="moderate",
    duration_weeks=8,
    monitoring_markers=[
        "fasting_glucose",
        "hba1c",
        "weight",
        "waist_circumference"
    ],
    safety_constraints=[
        "not a treatment prescription",
        "clinical review required if medication is involved",
        "stop or refer if red flags appear"
    ]
)

The engineering distinction is:

perturbation parameter != intervention action object

3. Systems biology often starts from a mechanism graph

A simplified systems biology model can be represented as a graph.

from dataclasses import dataclass
from typing import List

@dataclass
class BiologicalNode:
    node_id: str
    node_type: str  # gene, protein, metabolite, pathway, phenotype
    name: str

@dataclass
class BiologicalEdge:
    source: str
    target: str
    relation: str  # activates, inhibits, regulates, correlates_with
    evidence_strength: str

@dataclass
class MechanismGraph:
    nodes: List[BiologicalNode]
    edges: List[BiologicalEdge]

Example:

mechanism_graph = MechanismGraph(
    nodes=[
        BiologicalNode("n1", "pathway", "insulin_signaling"),
        BiologicalNode("n2", "phenotype", "glucose_variability"),
        BiologicalNode("n3", "phenotype", "fatigue")
    ],
    edges=[
        BiologicalEdge(
            source="n1",
            target="n2",
            relation="regulates",
            evidence_strength="moderate"
        ),
        BiologicalEdge(
            source="n2",
            target="n3",
            relation="associated_with",
            evidence_strength="low"
        )
    ]
)

This structure is valuable.

It helps represent:

mechanisms;
pathways;
regulatory relationships;
phenotype associations;
biological modules;
possible system-level interactions.

But it is not yet a full medical world model.

It still does not explicitly answer:

What is the current individual state?
What action may be taken?
How might state change after the action?
What evidence supports that transition?
What feedback window should be used?
How should the next cycle be updated if feedback differs from expectation?

That is where the world-model framing becomes useful.

4. A medical world model needs a State object

A medical world model starts with an individual state representation.

from dataclasses import dataclass
from typing import Dict, List, Optional

@dataclass
class PatientState:
    patient_id: str
    demographics: Dict
    clinical_markers: Dict
    lifestyle: Dict
    symptoms: List[str]
    medications: List[str]
    omics: Optional[Dict] = None
    wearable: Optional[Dict] = None
    mechanism_context: Optional[Dict] = None

Example:

state = PatientState(
    patient_id="P001",
    demographics={
        "age": 52,
        "sex": "unspecified"
    },
    clinical_markers={
        "bmi": 29.1,
        "fasting_glucose": 6.2,
        "hba1c": 6.0,
        "triglycerides": 2.1,
        "hdl_c": 0.95
    },
    lifestyle={
        "sleep_hours": 5.8,
        "exercise_frequency_per_week": 1,
        "diet_pattern": "high_refined_carbohydrate"
    },
    symptoms=[
        "fatigue",
        "post_meal_sleepiness"
    ],
    medications=[],
    mechanism_context={
        "possible_insulin_resistance": True,
        "possible_glucose_variability": True,
        "data_quality": "partial"
    }
)

The key principle is:

State should not be a data dump. It should be referenceable by actions, transitions, evidence, and feedback.

If a field cannot influence action selection, transition reasoning, safety filtering, or feedback updates, it may be noise rather than useful state.

5. Transition is not treatment-effect prediction

Developers may be tempted to write:

next_state = predict_next_state(state, action)

In medical settings, this is risky language.

It can sound like the model predicts individual treatment effects.

A safer and more precise abstraction is:

transition = estimate_transition_hypothesis(
    state=state,
    action=action,
    evidence=evidence_chain
)

That is: a transition hypothesis.

@dataclass
class TransitionHypothesis:
    expected_direction: Dict
    mechanism_rationale: List[str]
    uncertainty_level: str
    time_window_weeks: int
    assumptions: List[str]
    limitations: List[str]

Example:

transition = TransitionHypothesis(
    expected_direction={
        "fasting_glucose": "decrease_possible",
        "postprandial_glucose": "decrease_possible",
        "weight": "slight_decrease_possible",
        "energy_level": "may_improve"
    },
    mechanism_rationale=[
        "lower refined carbohydrate intake may reduce postprandial glucose excursions",
        "weight reduction may improve insulin sensitivity",
        "improved sleep may reduce metabolic stress"
    ],
    uncertainty_level="moderate",
    time_window_weeks=8,
    assumptions=[
        "adequate adherence",
        "no major medication change",
        "baseline data quality is acceptable"
    ],
    limitations=[
        "individual response may vary",
        "not a treatment effect prediction",
        "not a substitute for clinical judgment"
    ]
)

Notice the wording:

decrease_possible
may_improve
hypothesis
uncertainty
limitations

Not:

will decrease
will reverse
will cure

This distinction matters in medical AI.

A medical world model should not make deterministic treatment promises. It should produce mechanism-informed, evidence-bounded, uncertainty-aware transition hypotheses.

6. EvidenceChain: recommendations are not enough

If a system only outputs:

recommendation = "reduce refined carbohydrates and increase exercise"

it is not yet a medical world model.

A medical world model should explain:

why this action is proposed;
which mechanism it targets;
what evidence supports it;
what uncertainty remains;
where the safety boundary is.

A simple evidence chain can look like:

@dataclass
class EvidenceItem:
    source_type: str  # guideline, trial, mechanism, omics, individual_context
    description: str
    strength: str
    reference: str | None = None

@dataclass
class EvidenceChain:
    items: List[EvidenceItem]
    overall_strength: str
    uncertainty: str
    limitations: List[str]

Example:

evidence_chain = EvidenceChain(
    items=[
        EvidenceItem(
            source_type="mechanism",
            description="Reduced refined carbohydrate intake may reduce postprandial glucose excursions.",
            strength="moderate"
        ),
        EvidenceItem(
            source_type="individual_context",
            description="Current state includes high refined carbohydrate pattern and low exercise frequency.",
            strength="contextual"
        ),
        EvidenceItem(
            source_type="clinical_guideline",
            description="Lifestyle intervention is commonly recommended for metabolic risk management.",
            strength="high"
        )
    ],
    overall_strength="moderate",
    uncertainty="moderate",
    limitations=[
        "adherence is uncertain",
        "individual response may vary",
        "clinical review required when disease or medication is involved"
    ]
)

Engineering rule:

recommendation without evidence object = weak output
action + transition + evidence + feedback plan = stronger world-model output

7. Feedback: the system must update

A world model should not be a one-shot answer generator.

It should support feedback updates.

@dataclass
class FollowUpFeedback:
    patient_id: str
    action_id: str
    timepoint_weeks: int
    observed_markers: Dict
    adherence: Dict
    symptom_changes: Dict
    adverse_events: List[str]

Example:

feedback = FollowUpFeedback(
    patient_id="P001",
    action_id="nutrition_low_glycemic_8w",
    timepoint_weeks=8,
    observed_markers={
        "fasting_glucose": 5.8,
        "hba1c": 5.8,
        "weight_change_kg": -2.1,
        "waist_change_cm": -3.0
    },
    adherence={
        "nutrition": "medium",
        "exercise": "low",
        "sleep": "unchanged"
    },
    symptom_changes={
        "fatigue": "slightly_improved",
        "post_meal_sleepiness": "improved"
    },
    adverse_events=[]
)

Feedback then updates the record:

def update_world_model_state(
    previous_state: PatientState,
    action: InterventionAction,
    transition: TransitionHypothesis,
    feedback: FollowUpFeedback
):
    update_record = {
        "previous_state": previous_state,
        "action": action,
        "expected_transition": transition,
        "observed_feedback": feedback,
        "interpretation": None,
        "next_step": None
    }

    if feedback.adherence.get("nutrition") == "medium":
        update_record["interpretation"] = (
            "Partial improvement observed; adherence may limit effect size."
        )
        update_record["next_step"] = (
            "Review adherence barriers and consider adjusting action intensity."
        )
    else:
        update_record["interpretation"] = (
            "Observed feedback should be interpreted with caution."
        )
        update_record["next_step"] = (
            "Collect more context before updating the transition hypothesis."
        )

    return update_record

If the system cannot update from feedback, it is closer to a recommendation engine than a world model.

8. Causal boundaries: action-conditioned reasoning is not correlation

Once a system asks:

if action A, then what may happen?

it enters causal territory.

The transition should not be only correlation:

transition = correlate(state_features, future_outcomes)

The system should explicitly track causal assumptions and uncertainty:

@dataclass
class CausalAssumption:
    assumption_id: str
    description: str
    possible_confounders: List[str]
    applicable_population: str
    evidence_level: str
    uncertainty: str

Example:

causal_assumption = CausalAssumption(
    assumption_id="CA001",
    description=(
        "Reducing refined carbohydrate intake may reduce postprandial glucose "
        "excursions in individuals with diet-related glucose variability."
    ),
    possible_confounders=[
        "medication_change",
        "physical_activity_change",
        "sleep_change",
        "stress_change",
        "baseline_disease_status"
    ],
    applicable_population="health-management context with mild metabolic risk",
    evidence_level="moderate",
    uncertainty="individual_response_varies"
)

This does not mean every implementation must ship a full causal inference engine.

It means:

If the system outputs action-conditioned transitions, it must record causal assumptions, applicability, uncertainty, and limitations.

Otherwise, transition hypotheses can easily become correlation-based extrapolations.

9. SafetyGate: medical systems need boundaries before optimization

A medical world model should not simply optimize for the most promising action.

It must first apply safety boundaries.

@dataclass
class SafetyGateResult:
    passed: bool
    red_flags: List[str]
    contraindications: List[str]
    required_review: List[str]
    notes: List[str]

Example:

def run_safety_gate(
    state: PatientState,
    action: InterventionAction
) -> SafetyGateResult:
    red_flags = []
    contraindications = []
    required_review = []

    if state.clinical_markers.get("fasting_glucose", 0) > 13.9:
        red_flags.append("very_high_glucose_requires_clinical_evaluation")

    if "chest_pain" in state.symptoms:
        red_flags.append("chest_pain_requires_urgent_evaluation")

    if state.medications:
        required_review.append("medication_context_requires_clinician_review")

    passed = len(red_flags) == 0 and len(contraindications) == 0

    return SafetyGateResult(
        passed=passed,
        red_flags=red_flags,
        contraindications=contraindications,
        required_review=required_review,
        notes=[
            "not medical advice",
            "not a validated treatment planning system",
            "human review required in clinical context"
        ]
    )

Principle:

No safety gate, no medical world-model deployment.

10. AuditLog: every transition should leave a trace

A medical world model should be able to answer:

What was the state at the time?
Why was this action proposed?
What was the transition hypothesis?
What evidence supported it?
Who reviewed it?
Did feedback match the expectation?
If not, how was the next cycle updated?

A minimal audit log:

@dataclass
class AuditLog:
    record_id: str
    patient_id: str
    state_snapshot_id: str
    action_id: str
    transition_id: str
    evidence_chain_id: str
    safety_gate_id: str
    reviewer: str
    decision: str
    timestamp: str

Example:

audit_log = AuditLog(
    record_id="AUDIT_20260521_001",
    patient_id="P001",
    state_snapshot_id="STATE_20260521",
    action_id="nutrition_low_glycemic_8w",
    transition_id="TRANSITION_20260521_001",
    evidence_chain_id="EVIDENCE_20260521_001",
    safety_gate_id="SAFETY_20260521_001",
    reviewer="human_expert",
    decision="approved_for_health_management_context",
    timestamp="2026-05-21T20:00:00+08:00"
)

The goal is not to generate a better-sounding answer.

The goal is to make the reasoning process traceable, auditable, feedback-driven, and reviewable.

11. A minimal medical world-model loop

Putting the objects together:

def medical_world_model_loop(patient_id: str):
    # 1. Observe current state
    state = observe_patient_state(patient_id)

    # 2. Retrieve mechanism context
    mechanism_context = retrieve_mechanism_context(state)

    # 3. Generate candidate actions
    candidate_actions = generate_candidate_actions(
        state=state,
        mechanism_context=mechanism_context
    )

    transition_candidates = []

    for action in candidate_actions:
        # 4. Safety gate first
        safety = run_safety_gate(state, action)

        if not safety.passed:
            continue

        # 5. Build evidence chain
        evidence = build_evidence_chain(
            state=state,
            action=action,
            mechanism_context=mechanism_context
        )

        # 6. Estimate transition hypothesis
        transition = estimate_transition_hypothesis(
            state=state,
            action=action,
            evidence=evidence
        )

        transition_candidates.append({
            "action": action,
            "transition": transition,
            "evidence": evidence,
            "safety": safety
        })

    # 7. Human-in-the-loop review
    selected = human_expert_review(transition_candidates)

    # 8. Collect follow-up feedback
    feedback = collect_follow_up_feedback(
        patient_id=patient_id,
        action_id=selected["action"].action_id,
        time_window_weeks=selected["transition"].time_window_weeks
    )

    # 9. Update model state
    updated_record = update_world_model_state(
        previous_state=state,
        action=selected["action"],
        transition=selected["transition"],
        feedback=feedback
    )

    # 10. Write audit log
    audit_log = write_audit_log(
        state=state,
        selected=selected,
        feedback=feedback,
        updated_record=updated_record
    )

    return {
        "updated_record": updated_record,
        "audit_log": audit_log
    }

The workflow order matters:

state
  -> mechanism context
  -> candidate action
  -> safety gate
  -> evidence chain
  -> transition hypothesis
  -> human review
  -> feedback
  -> update
  -> audit log

That is the engineering difference between a mechanism graph and a medical world-model loop.

12. JSON example: one transition record

A simplified transition record may look like:

{
  "state": {
    "patient_id": "P001",
    "state_snapshot_id": "STATE_20260521",
    "clinical_markers": {
      "bmi": 29.1,
      "fasting_glucose": 6.2,
      "hba1c": 6.0,
      "triglycerides": 2.1
    },
    "lifestyle": {
      "sleep_hours": 5.8,
      "exercise_frequency_per_week": 1,
      "diet_pattern": "high_refined_carbohydrate"
    },
    "mechanism_context": {
      "possible_insulin_resistance": true,
      "possible_glucose_variability": true,
      "data_quality": "partial"
    }
  },
  "action": {
    "action_id": "nutrition_low_glycemic_8w",
    "category": "nutrition",
    "duration_weeks": 8,
    "target_mechanisms": [
      "postprandial_glucose_variability",
      "insulin_resistance"
    ],
    "monitoring_markers": [
      "fasting_glucose",
      "hba1c",
      "weight",
      "waist_circumference"
    ]
  },
  "transition_hypothesis": {
    "expected_direction": {
      "fasting_glucose": "decrease_possible",
      "postprandial_glucose": "decrease_possible",
      "weight": "slight_decrease_possible"
    },
    "uncertainty_level": "moderate",
    "time_window_weeks": 8,
    "limitations": [
      "individual_response_varies",
      "not_a_treatment_effect_prediction"
    ]
  },
  "evidence_chain": {
    "overall_strength": "moderate",
    "items": [
      {
        "source_type": "mechanism",
        "description": "Reduced refined carbohydrate intake may reduce postprandial glucose excursions."
      },
      {
        "source_type": "individual_context",
        "description": "Current lifestyle pattern includes high refined carbohydrate intake."
      }
    ]
  },
  "safety_gate": {
    "passed": true,
    "red_flags": [],
    "notes": [
      "not_medical_advice",
      "human_review_required_in_clinical_context"
    ]
  },
  "feedback_plan": {
    "timepoint_weeks": 8,
    "metrics": [
      "fasting_glucose",
      "hba1c",
      "weight",
      "waist_circumference",
      "symptom_score"
    ]
  }
}

This schema is not the point.

The point is that the reasoning becomes structured.

13. Developer principles

Principle 1: Do not start with a chatbot

Avoid starting with:

answer = llm.chat(user_question)

Start with objects:

state_schema = define_state_schema()
action_schema = define_action_schema()
transition_schema = define_transition_schema()
evidence_schema = define_evidence_schema()
feedback_schema = define_feedback_schema()

Principle 2: Do not frame transition as treatment-effect prediction

Avoid:

effect = predict_treatment_effect(state, action)

Prefer:

transition = estimate_transition_hypothesis(state, action, evidence)

Principle 3: A mechanism graph is not the full world model

mechanism_graph = build_mechanism_graph(omics_data)

This is useful, but not enough.

You still need:

action = define_intervention_action()
transition = estimate_transition_hypothesis(state, action, evidence)
feedback = collect_follow_up_feedback()

Principle 4: Evidence must be a first-class object

Avoid:

recommendation = generate_recommendation(state)

Prefer:

output = {
    "state": state,
    "action": action,
    "transition_hypothesis": transition,
    "evidence_chain": evidence_chain,
    "safety_gate": safety_gate,
    "feedback_plan": feedback_plan
}

Principle 5: Human-in-the-loop is core

A medical world model should not be designed as an automatic treatment system.

decision = human_expert_review(model_output)

This should be part of the architecture, not an afterthought.

Principle 6: No feedback, no strong world model

If the system cannot update:

updated_state = update_state(previous_state, action, feedback)

it is closer to a one-shot recommendation system than a medical world model.

14. A steerable medical world model

In this context, SteeraMed can be understood as a steerable biomedical world-model framework.

Its engineering focus is not automatic control of the human body.

It is the organization of these objects:

State
Action
Transition Hypothesis
Evidence Chain
Safety Gate
Human Review
Feedback
Audit Log

A simplified interface:

class SteerableMedicalWorldModel:
    def observe_state(self, patient_id):
        pass

    def generate_actions(self, state):
        pass

    def run_safety_gate(self, state, action):
        pass

    def build_evidence_chain(self, state, action):
        pass

    def estimate_transition(self, state, action, evidence):
        pass

    def request_human_review(self, candidates):
        pass

    def collect_feedback(self, selected_action):
        pass

    def update_model(self, state, action, feedback):
        pass

    def write_audit_log(self, record):
        pass

This is very different from a medical chatbot.

A chatbot can generate a plausible answer.

A medical world model should preserve the state, action, evidence, transition, feedback, and audit trail behind the answer.

15. Summary: mechanism layer vs action-simulation layer

Systems biology is essential.

It helps us understand biological networks, pathways, mechanisms, dynamics, and system-level regulation.

But from an engineering perspective, a systems biology model is usually not yet a complete medical world model.

A medical world model connects:

individual state
intervention action
mechanism-informed evidence
transition hypothesis
safety gate
human review
longitudinal feedback
audit log

The relationship can be summarized as:

Systems biology:
    mechanism understanding

Medical world model:
    mechanism-informed action simulation

Steerable medical world model:
    goal-directed, evidence-bounded, feedback-calibrated intervention reasoning

So the distinction is not about replacing systems biology.

It is about extending mechanism understanding into action-conditioned, evidence-bounded, feedback-calibrated reasoning.

That is why medical AI still needs medical world models, even in a field where systems biology is already powerful.

References

Kitano, H. Systems Biology: A Brief Overview. Science, 2002. https://doi.org/10.1126/science.1069492
Kitano, H. Computational systems biology. Nature, 2002. https://doi.org/10.1038/nature01254
Ideker, T., Galitski, T., & Hood, L. A new approach to decoding life: systems biology. Annual Review of Genomics and Human Genetics, 2001. https://doi.org/10.1146/annurev.genom.2.1.343
Barabási, A.-L., Gulbahce, N., & Loscalzo, J. Network medicine: a network-based approach to human disease. Nature Reviews Genetics, 2011. https://doi.org/10.1038/nrg2918
Noble, D. The Music of Life: Biology Beyond Genes. Oxford University Press, 2006.
Ha, D., & Schmidhuber, J. Recurrent World Models Facilitate Policy Evolution. NeurIPS, 2018. https://arxiv.org/abs/1803.10122
LeCun, Y. A Path Towards Autonomous Machine Intelligence. OpenReview, 2022. https://openreview.net/forum?id=BZ5a1r-kVsf
Pearl, J., & Mackenzie, D. The Book of Why: The New Science of Cause and Effect. Basic Books, 2018.
Katsoulakis, E., Wang, Q., Wu, H., et al. Digital twins for health: a scoping review. npj Digital Medicine, 2024. https://doi.org/10.1038/s41746-024-01073-0
Yang, Y., Wang, Z.-Y., Liu, Q., Sun, S., Wang, K., Chellappa, R., Zhou, Z., Yuille, A., Zhu, L., Zhang, Y.-D., & Chen, J. Medical World Model: Generative Simulation of Tumor Evolution for Treatment Planning. arXiv:2506.02327, 2025. https://arxiv.org/abs/2506.02327
Xiong, J. World Models for Biomedicine: A Steerability Framework. Preprints.org, 2026. https://doi.org/10.20944/preprints202605.0366.v1
SteeraMed project: https://SteeraMed.com
Steerable World project: https://steerable.world

After AI Healthcare, Medical World Models May Be the Next Life-Science AI Platform

JXIONG — Thu, 21 May 2026 10:32:55 +0000

Subtitle: A system-design view of moving from risk prediction to intervention simulation**

Over the last decade, most AI healthcare narratives have been about helping machines see disease.

Computer vision systems detect lesions in medical images. Risk models estimate the probability of cardiovascular events, diabetes, readmission, or poor outcomes. Large language models summarize clinical notes, explain lab reports, and assist with medical text workflows.

These capabilities matter.

But most of them still answer one of two questions:

What is the current state?
What might happen in the future?

Over the last few years, AI drug discovery has become one of the most visible frontiers in life-science AI. AI is now being used for target discovery, molecule generation, protein modeling, virtual screening, and trial optimization.

That is a major shift: AI is no longer only helping us identify disease; it is also helping us discover molecules.

But there may be another layer ahead.

The next life-science AI platform may not be only about identifying disease or discovering molecules. It may be about building systems that can represent an individual's biological state, encode possible interventions, simulate state-transition hypotheses, track evidence, and update decisions through longitudinal feedback.

That is the idea behind a medical world model.

A medical world model does not simply ask:

What is the patient's risk?

It asks:

If we take this action, how might the patient's state change?
Why does the model believe that transition is plausible?
What evidence supports it?
What feedback should update the next decision?

This article explains that idea from a system-design perspective.

1. Healthcare AI mostly started with recognition and prediction

Many healthcare AI systems can be simplified into three categories:

Recognition
- Is this image abnormal?
- Is there a lesion?
- Is this ECG pattern suspicious?
Classification
- Which subtype does this case belong to?
- Which risk group is this patient in?
Prediction
- What is the probability of a future event?
- How likely is readmission?
- What is the estimated disease risk?

A typical medical prediction model looks like this:

risk = predict_risk(patient_state)

For example:

patient_state = {
    "age": 52,
    "bmi": 29.1,
    "fasting_glucose": 6.2,
    "hba1c": 6.0,
    "blood_pressure": "138/86",
    "family_history": ["type_2_diabetes"],
    "sleep_duration": 5.8
}

risk = predict_diabetes_risk(patient_state)

The output might be:

{
  "risk_level": "high",
  "estimated_5y_risk": 0.32
}

This answers:

How high is the future risk?

That is useful.

But real medical and health-management decisions do not stop there.

The next questions are usually:

What should be done first?
Should nutrition, exercise, sleep, medication review, or follow-up be prioritized?
Which intervention best matches the current mechanism hypothesis?
Which markers should be monitored?
If the expected change does not occur, was the action wrong, the mechanism wrong, or the feedback window wrong?

At that point, the system needs something most prediction models do not explicitly represent:

Action

2. What does a medical world model model?

A medical world model is not a larger medical chatbot.

It is not an automatic treatment generator.

It is better understood as an auditable inference architecture built around five objects:

State       The current individual state
Action      A defined intervention or decision option
Transition  A hypothesis about how state may change after action
Evidence    The evidence chain supporting the hypothesis
Feedback    Real-world follow-up used to update the model

A prediction model often looks like:

state -> outcome

A medical world model looks more like:

state + action + evidence -> transition hypothesis -> feedback update

In other words:

Prediction model:
    What may happen?

Medical world model:
    What may happen if we act?

This is the shift from risk prediction to intervention simulation.

3. State: represent the individual before reasoning about action

The first step is not training a bigger model.

The first step is defining the state.

A simplified PatientState object might look like this:

from dataclasses import dataclass
from typing import Dict, List, Optional

@dataclass
class PatientState:
    demographics: Dict
    clinical_markers: Dict
    symptoms: List[str]
    lifestyle: Dict
    medications: List[str]
    history: Dict
    omics: Optional[Dict] = None
    wearable: Optional[Dict] = None

Example:

patient_state = PatientState(
    demographics={
        "age": 52,
        "sex": "unspecified"
    },
    clinical_markers={
        "bmi": 29.1,
        "fasting_glucose": 6.2,
        "hba1c": 6.0,
        "triglycerides": 2.1,
        "hdl_c": 0.95,
        "blood_pressure": "138/86"
    },
    symptoms=[
        "fatigue",
        "post_meal_sleepiness"
    ],
    lifestyle={
        "sleep_hours": 5.8,
        "exercise_frequency_per_week": 1,
        "diet_pattern": "high_refined_carbohydrate",
        "stress_level": "high"
    },
    medications=[],
    history={
        "family_history": ["type_2_diabetes"],
        "previous_diagnosis": []
    }
)

The goal is not to add endless fields.

The goal is to create a state representation that can support:

action selection;
evidence retrieval;
transition estimation;
safety checking;
feedback updates.

A state that cannot be referenced by actions or updated through feedback is not very useful for a world-model system.

4. Action: make interventions computable

Prediction models do not necessarily need actions.

Medical world models do.

The phrase "improve lifestyle" is not a good action object. It is too vague to execute, track, audit, or update.

A better approach is to encode interventions as structured objects:

@dataclass
class InterventionAction:
    action_id: str
    category: str
    description: str
    target_mechanism: List[str]
    intensity: str
    duration_weeks: int
    monitoring_markers: List[str]
    safety_notes: List[str]

Example:

action = InterventionAction(
    action_id="nutrition_low_glycemic_8w",
    category="nutrition",
    description="8-week low-glycemic dietary adjustment with reduced refined carbohydrates",
    target_mechanism=[
        "postprandial_glucose_variability",
        "insulin_resistance",
        "weight_management"
    ],
    intensity="moderate",
    duration_weeks=8,
    monitoring_markers=[
        "fasting_glucose",
        "hba1c",
        "weight",
        "waist_circumference",
        "postprandial_glucose"
    ],
    safety_notes=[
        "not a medical prescription",
        "review with clinician if diabetes medication is used",
        "monitor hypoglycemia risk when relevant"
    ]
)

This matters because a medical world model should not merely generate recommendations.

It should make each action:

describable;
executable;
trackable;
auditable;
reviewable;
feedback-compatible.

5. Transition: a hypothesis, not a treatment-effect promise

In ordinary engineering language, you may be tempted to write:

next_state = model.predict_next_state(state, action)

In medicine, that can be misleading.

It sounds like the system is predicting individual treatment effects.

A safer and more accurate name is:

transition_hypothesis = estimate_transition_tendency(state, action)

A transition object might look like:

@dataclass
class TransitionHypothesis:
    expected_direction: Dict
    mechanism_rationale: List[str]
    uncertainty_level: str
    time_window_weeks: int
    assumptions: List[str]

Example:

transition = TransitionHypothesis(
    expected_direction={
        "fasting_glucose": "decrease_possible",
        "postprandial_glucose": "decrease_possible",
        "weight": "slight_decrease_possible",
        "energy_level": "may_improve"
    },
    mechanism_rationale=[
        "lower refined carbohydrate intake may reduce postprandial glucose excursion",
        "weight reduction may improve insulin sensitivity",
        "improved dietary pattern may reduce metabolic stress"
    ],
    uncertainty_level="moderate",
    time_window_weeks=8,
    assumptions=[
        "adequate adherence",
        "no major medication change",
        "baseline data quality is acceptable",
        "no unrecognized endocrine disorder"
    ]
)

Notice what this does not say:

will cure
will reverse
will normalize
will improve with certainty

Instead, it says:

decrease_possible
may_improve
transition tendency

That distinction is essential.

A medical world model should generate mechanism-constrained transition hypotheses, not deterministic treatment promises.

6. Evidence: every transition needs an evidence chain

A transition without evidence is just a generated suggestion.

A medical world model needs an evidence object.

@dataclass
class EvidenceItem:
    source_type: str
    description: str
    strength: str
    url_or_reference: Optional[str] = None

@dataclass
class EvidenceChain:
    items: List[EvidenceItem]
    overall_strength: str
    limitations: List[str]

Example:

evidence_chain = EvidenceChain(
    items=[
        EvidenceItem(
            source_type="clinical_guideline",
            description="Lifestyle modification is commonly recommended for metabolic risk management.",
            strength="high"
        ),
        EvidenceItem(
            source_type="mechanistic_evidence",
            description="Reduced refined carbohydrate intake may lower postprandial glucose excursions.",
            strength="moderate"
        ),
        EvidenceItem(
            source_type="individual_context",
            description="Patient reports high refined carbohydrate intake and low exercise frequency.",
            strength="contextual"
        )
    ],
    overall_strength="moderate",
    limitations=[
        "individual response may vary",
        "adherence is uncertain",
        "not a substitute for clinical evaluation"
    ]
)

The evidence object should help answer:

Where does the reasoning come from?
How strong is the evidence?
What are the assumptions?
What is the uncertainty?
What are the clinical or safety boundaries?

Without this layer, a medical world model risks becoming a black-box recommendation engine.

7. Feedback: the model must update over time

A world model is not a one-shot answer generator.

It must support feedback.

@dataclass
class FollowUpFeedback:
    timepoint_weeks: int
    observed_markers: Dict
    adherence: Dict
    symptoms_change: Dict
    adverse_events: List[str]

Example:

feedback = FollowUpFeedback(
    timepoint_weeks=8,
    observed_markers={
        "fasting_glucose": 5.8,
        "hba1c": 5.8,
        "weight": -2.1,
        "waist_circumference": -3.0
    },
    adherence={
        "diet": "medium",
        "exercise": "low",
        "sleep": "unchanged"
    },
    symptoms_change={
        "fatigue": "slightly_improved",
        "post_meal_sleepiness": "improved"
    },
    adverse_events=[]
)

Then update the record:

def update_state_with_feedback(
    previous_state: PatientState,
    action: InterventionAction,
    transition: TransitionHypothesis,
    feedback: FollowUpFeedback
):
    audit_log = {
        "previous_state": previous_state,
        "action": action,
        "expected_transition": transition,
        "observed_feedback": feedback,
        "interpretation": None,
        "next_step": None
    }

    if feedback.adherence["diet"] == "medium":
        audit_log["interpretation"] = (
            "Partial improvement observed; adherence may limit effect size."
        )
        audit_log["next_step"] = (
            "Review action intensity and adherence barriers."
        )
    else:
        audit_log["interpretation"] = (
            "Feedback should be interpreted with caution."
        )
        audit_log["next_step"] = (
            "Collect more context before updating intervention plan."
        )

    return audit_log

The key loop is:

observe -> act -> simulate -> monitor -> update

From a platform perspective, this is important.

The next generation of medical AI may not be a single-use diagnostic tool. It may be a longitudinal feedback platform.

8. A minimal medical world-model workflow

A minimal workflow could look like this:

def medical_world_model_loop(patient_id: str):
    # 1. Observe state
    state = observe_patient_state(patient_id)

    # 2. Generate candidate actions
    candidate_actions = generate_candidate_actions(state)

    # 3. Safety filter
    safe_actions = []
    for action in candidate_actions:
        if pass_safety_gate(state, action):
            safe_actions.append(action)

    # 4. Estimate transitions
    transition_candidates = []
    for action in safe_actions:
        transition = estimate_transition_tendency(state, action)
        evidence = build_evidence_chain(state, action, transition)

        transition_candidates.append({
            "action": action,
            "transition": transition,
            "evidence": evidence
        })

    # 5. Human-in-the-loop review
    selected_action = clinician_or_expert_review(transition_candidates)

    # 6. Execute and monitor
    feedback = collect_follow_up_feedback(patient_id, selected_action)

    # 7. Update state and audit log
    updated_record = update_state_with_feedback(
        previous_state=state,
        action=selected_action,
        transition=selected_action["transition"],
        feedback=feedback
    )

    return updated_record

The most important line is this:

selected_action = clinician_or_expert_review(transition_candidates)

A medical world model should not bypass professional review.

Its safer positioning is:

hypothesis generation + decision support + audit trail

Not:

automatic diagnosis or treatment

9. Safety gate: boundaries must come before optimization

In medical systems, safety should not be an afterthought.

def pass_safety_gate(state: PatientState, action: InterventionAction) -> bool:
    # Example checks only. Not medical advice.
    contraindications = detect_contraindications(state, action)
    medication_conflicts = check_medication_conflicts(state, action)
    red_flags = detect_red_flags(state)

    if red_flags:
        return False

    if contraindications:
        return False

    if medication_conflicts:
        return False

    return True

Example:

def detect_red_flags(state: PatientState) -> List[str]:
    red_flags = []

    if state.clinical_markers.get("fasting_glucose", 0) > 13.9:
        red_flags.append("very_high_glucose_requires_clinical_evaluation")

    if "chest_pain" in state.symptoms:
        red_flags.append("chest_pain_requires_urgent_evaluation")

    return red_flags

The design principle is simple:

A medical AI system should not become more autonomous faster than it becomes auditable.

10. Audit logs are not optional

A medical world model should leave an audit trail for every transition hypothesis.

@dataclass
class AuditLog:
    patient_id: str
    state_snapshot_id: str
    action_id: str
    transition_id: str
    evidence_chain_id: str
    reviewer: str
    decision: str
    uncertainty_level: str
    safety_notes: List[str]
    timestamp: str

Example:

audit_log = AuditLog(
    patient_id="P001",
    state_snapshot_id="S20260521",
    action_id="nutrition_low_glycemic_8w",
    transition_id="T20260521_001",
    evidence_chain_id="E20260521_001",
    reviewer="human_expert",
    decision="approved_for_health_management_context",
    uncertainty_level="moderate",
    safety_notes=[
        "not medical diagnosis",
        "not treatment prescription",
        "clinical review required if symptoms worsen"
    ],
    timestamp="2026-05-21T17:00:00+08:00"
)

Without audit logs, the system cannot answer:

Why was this action proposed?
What evidence supported it?
Which assumptions later failed?
Which feedback changed the next decision?
Where should responsibility and review occur?

This is where medical world models differ from ordinary generative AI applications.

11. Steerable world models: not control, but direction and feedback

A regular world model can simulate possible futures.

Medicine needs more than simulation.

It needs a way to define objectives, actions, boundaries, feedback metrics, and stop conditions.

That is the idea behind a steerable world model.

Steerable does not mean controlling the human body.

It means making the intervention loop explicit:

@dataclass
class SteeringInterface:
    objective: Dict
    allowed_actions: List[InterventionAction]
    safety_constraints: List[str]
    feedback_metrics: List[str]
    stop_conditions: List[str]

Example:

steering = SteeringInterface(
    objective={
        "primary": "improve_metabolic_resilience",
        "secondary": ["reduce_glucose_variability", "improve_energy_level"]
    },
    allowed_actions=[
        "nutrition_adjustment",
        "exercise_adjustment",
        "sleep_management",
        "clinical_referral_when_needed"
    ],
    safety_constraints=[
        "no medication change without clinician",
        "stop if red flags appear",
        "avoid unsupported intervention claims"
    ],
    feedback_metrics=[
        "fasting_glucose",
        "postprandial_glucose",
        "weight",
        "waist_circumference",
        "symptom_score"
    ],
    stop_conditions=[
        "adverse_event",
        "red_flag_symptom",
        "data_quality_insufficient"
    ]
)

For medical AI, steerability means:

objective;
action;
boundary;
evidence;
feedback;
stop condition;
human review.

Not autonomous control.

12. Why investors should pay attention to medical world models

The investment relevance is not that "medical world model" is a new buzzword.

The relevance is that it may connect several currently fragmented life-science AI markets.

1. It extends healthcare AI

Healthcare AI started with recognition, classification, and prediction.

Medical world models extend that into intervention simulation.

2. It complements AI drug discovery

AI drug discovery focuses on targets, molecules, and development workflows.

Medical world models focus on what happens when interventions meet individual states.

That can include drugs, but also nutrition, exercise, sleep, behavioral interventions, monitoring, and long-term care pathways.

3. It provides a framework for precision medicine

Precision medicine needs individualized state representation and decision logic.

Medical world models provide a structure for state, action, transition, evidence, and feedback.

4. It fits longevity medicine

Longevity medicine is not a one-time diagnosis.

It is longitudinal state management.

That makes it naturally aligned with state-action-transition-feedback loops.

5. It may become a platform layer

The platform opportunity is not a single model output.

It is a longitudinal infrastructure for:

state representation;
intervention encoding;
evidence tracking;
safety filtering;
expert review;
feedback collection;
model calibration.

That is why medical world models may represent a future life-science AI platform category rather than just another AI tool.

13. Why longevity medicine is a natural early use case

Longevity medicine deals with long-term state management rather than single-point diagnosis.

It involves:

multi-system aging;
metabolic and immune changes;
chronic low-grade inflammation;
sleep, stress, movement, and nutrition;
individual differences;
combined interventions;
periodic retesting;
N-of-1 feedback.

This is not just a classification problem.

It is a longitudinal loop:

while health_management_active:
    state = observe_longitudinal_state(user)
    actions = generate_intervention_candidates(state)
    transitions = estimate_transition_tendencies(state, actions)
    reviewed_plan = human_review(transitions)
    feedback = collect_longitudinal_feedback(reviewed_plan)
    update_model_state(state, reviewed_plan, feedback)

In system terms:

longevity medicine = longitudinal state-action-transition-feedback problem

That is why longevity tech, precision health, and functional medicine may become early application environments for medical world models.

14. A compact JSON representation

Here is a simplified JSON representation of a medical world-model record:

{
  "patient_state": {
    "state_id": "S20260521",
    "clinical_markers": {
      "bmi": 29.1,
      "fasting_glucose": 6.2,
      "hba1c": 6.0,
      "triglycerides": 2.1
    },
    "lifestyle": {
      "sleep_hours": 5.8,
      "exercise_frequency_per_week": 1,
      "diet_pattern": "high_refined_carbohydrate"
    },
    "risk_context": [
      "family_history_type_2_diabetes",
      "possible_insulin_resistance"
    ]
  },
  "candidate_action": {
    "action_id": "nutrition_low_glycemic_8w",
    "category": "nutrition",
    "duration_weeks": 8,
    "target_mechanism": [
      "postprandial_glucose_variability",
      "insulin_resistance"
    ],
    "monitoring_markers": [
      "fasting_glucose",
      "hba1c",
      "weight",
      "waist_circumference"
    ]
  },
  "transition_hypothesis": {
    "expected_direction": {
      "fasting_glucose": "decrease_possible",
      "postprandial_glucose": "decrease_possible",
      "weight": "slight_decrease_possible"
    },
    "uncertainty_level": "moderate",
    "time_window_weeks": 8
  },
  "evidence_chain": {
    "overall_strength": "moderate",
    "limitations": [
      "individual_response_varies",
      "adherence_uncertain",
      "not_a_treatment_prescription"
    ]
  },
  "safety_gate": {
    "requires_clinician_review": false,
    "red_flags": [],
    "notes": [
      "health_management_context_only",
      "not_medical_diagnosis"
    ]
  },
  "feedback_plan": {
    "timepoint_weeks": 8,
    "metrics": [
      "fasting_glucose",
      "hba1c",
      "weight",
      "waist_circumference",
      "symptom_score"
    ]
  }
}

The point is not this exact schema.

The point is that a medical world model decomposes reasoning into inspectable objects.

15. Developer principles

Principle 1: Do not start with a chatbot

A medical world model should not begin with:

answer = llm.chat(user_question)

It should begin with schemas:

state = define_state_schema()
action = define_action_schema()
transition = define_transition_schema()
evidence = define_evidence_schema()
feedback = define_feedback_schema()

Principle 2: Do not frame transition as treatment-effect prediction

Avoid:

effect = predict_treatment_effect(state, action)

Prefer:

hypothesis = estimate_transition_tendency(state, action, evidence)

Principle 3: Evidence must be a first-class object

Avoid:

recommendation = generate_recommendation(state)

Prefer:

recommendation = {
    "action": action,
    "transition_hypothesis": transition,
    "evidence_chain": evidence,
    "uncertainty": uncertainty,
    "safety_notes": safety_notes
}

Principle 4: Human-in-the-loop should be core

decision = human_expert_review(model_output)

This should be part of the design, not an afterthought.

Principle 5: Feedback update is the product moat

If there is no feedback update, the system is not really a world model.

model_state = update_with_feedback(model_state, observed_feedback)

16. From tool to infrastructure

Healthcare AI's first wave helped machines see disease.

AI drug discovery helped machines search molecular space.

Medical world models may help machines reason about interventions under uncertainty.

From an engineering perspective, the architecture is:

State
  + Action
  + Evidence
  -> Transition Hypothesis
  -> Feedback
  -> Calibration

The value is not "automatic treatment."

The value is making medical reasoning:

representable;
auditable;
traceable;
feedback-driven;
calibratable;
reviewable by human experts.

For longevity medicine, precision health, functional medicine, and long-term health-management platforms, this architecture may be especially important.

Those fields do not need one-shot predictions.

They need longitudinal state-action-transition-feedback loops.

If healthcare AI's first value was to see disease, and AI drug discovery's value is to discover molecules, then medical world models may define the next stage:

simulate interventions,
track feedback,
and continuously calibrate individual biological states.

That is why medical world models may become a next-generation life-science AI platform category.

References

Ha, D., & Schmidhuber, J. Recurrent World Models Facilitate Policy Evolution. Advances in Neural Information Processing Systems 31, 2018. https://arxiv.org/abs/1803.10122
LeCun, Y. A Path Towards Autonomous Machine Intelligence. OpenReview, 2022. https://openreview.net/forum?id=BZ5a1r-kVsf
Yang, Y., Wang, Z.-Y., Liu, Q., Sun, S., Wang, K., Chellappa, R., Zhou, Z., Yuille, A., Zhu, L., Zhang, Y.-D., & Chen, J. Medical World Model: Generative Simulation of Tumor Evolution for Treatment Planning. arXiv:2506.02327, 2025. https://arxiv.org/abs/2506.02327
Qazi, M. A., Nadeem, M., & Yaqub, M. Beyond Generative AI: World Models for Clinical Prediction, Counterfactuals, and Planning. arXiv:2511.16333, 2025. https://arxiv.org/abs/2511.16333
Katsoulakis, E., Wang, Q., Wu, H., et al. Digital twins for health: a scoping review. npj Digital Medicine, 7, 77, 2024. https://doi.org/10.1038/s41746-024-01073-0
Pearl, J., & Mackenzie, D. The Book of Why: The New Science of Cause and Effect. Basic Books, 2018.
Xiong, J. World Models for Biomedicine: A Steerability Framework. Preprints.org, 2026. https://doi.org/10.20944/preprints202605.0366.v1
Steerable World project: https://steerable.world

What Is a World Model, and Why Is It More Than Prediction?

JXIONG — Thu, 21 May 2026 00:38:09 +0000

Most medical AI systems today are still designed as prediction systems.

A typical pipeline looks like this:

data = collect_patient_data()
features = extract_features(data)
risk = model.predict(features)
return risk

This can be useful.

A prediction model can answer questions such as:

What is the estimated risk of a disease?
Does an image contain an abnormal finding?
Which risk group does this person belong to?
What is the probability of a future clinical event?

But a world model asks a different question.

A prediction model asks:

future = predict(state)

A world model asks:

next_state = simulate_transition(state, action)

In other words:

Prediction asks: what may happen next?

A world model asks: what may happen if we take a specific action?

That difference matters a lot in medicine.

Medicine is not only about recognizing risk. It is also about deciding what to do under uncertainty.

The Minimal Structure of a World Model

A simplified world model can be described with five objects:

State       What is the system like now?
Action      What can be done?
Transition  How may the system change after the action?
Objective   What direction are we trying to move toward?
Feedback    What did we observe after the action?

A very small abstraction could look like this:

class WorldModel:
    def observe_state(self, system):
        pass

    def define_action(self, action_input):
        pass

    def simulate_transition(self, state, action):
        pass

    def collect_feedback(self, system, action):
        pass

    def update(self, state, action, feedback):
        pass

This is already different from a simple predictive model.

A predictive model can work without an explicit action object.

A world model cannot.

If there is no action, there is no action-conditioned transition.

Why Medical AI Needs More Than Risk Prediction

Many medical AI systems are good at recognition and prediction:

image classification;
risk scoring;
disease detection;
prognosis estimation;
anomaly detection;
population stratification.

But many real medical and health-management problems are not just classification problems.

They are action problems.

For example:

Should an intervention be considered?
Which variable should be monitored first?
What evidence supports the expected change?
What feedback should be collected?
If the result is different from expected, what should be updated?

These are not just prediction questions. They require a system to reason about state, action, transition, evidence, and feedback.

That is where the idea of a medical world model becomes useful.

A Prediction Model vs. a Medical World Model

A prediction model may look like this:

class PredictionModel:
    def predict_risk(self, patient_data):
        features = self.extract_features(patient_data)
        risk_score = self.model.predict(features)
        return risk_score

It maps data to a risk score.

A medical world model needs a different structure:

class MedicalWorldModel:
    def simulate_transition(self, state, action, evidence):
        if not self.safety_gate(state, action):
            return {
                "status": "blocked",
                "reason": "Safety gate failed. Human review required."
            }

        transition = self.transition_model.estimate(
            state=state,
            action=action,
            evidence=evidence
        )

        return {
            "status": "hypothesis_generated",
            "state": state,
            "action": action,
            "expected_transition": transition,
            "evidence": evidence,
            "disclaimer": "Hypothesis-generating only. Not medical advice."
        }

The output is not a treatment recommendation.

It is a transition hypothesis.

That distinction is critical.

State: Do Not Reduce a Person to a Risk Score

A risk score may be useful:

{
  "cardiovascular_risk": 0.23
}

But it is not enough for a world model.

A world model needs a richer representation of state.

Example schema:

{
  "subject_id": "anonymous_001",
  "timestamp": "2026-05-20",
  "metabolic_state": {
    "fasting_glucose": 5.6,
    "hba1c": 5.7,
    "fasting_insulin": 12.4,
    "triglycerides": 1.8
  },
  "inflammation_state": {
    "hs_crp": 2.1
  },
  "lifestyle_state": {
    "sleep_duration": 6.2,
    "weekly_exercise_minutes": 90,
    "diet_pattern": "high_refined_carbohydrate"
  },
  "risk_context": {
    "family_history": ["type_2_diabetes"],
    "medications": [],
    "known_conditions": []
  }
}

This is only an illustrative schema, not a clinical standard.

In a real system, every state variable should have:

source;
unit;
timestamp;
measurement context;
missing-value handling;
data-quality metadata;
uncertainty annotation.

A simple Python representation might look like this:

from dataclasses import dataclass
from typing import Dict, Any, List

@dataclass
class HealthState:
    subject_id: str
    timestamp: str
    biomarkers: Dict[str, Any]
    lifestyle: Dict[str, Any]
    symptoms: Dict[str, Any]
    medications: List[str]
    context: Dict[str, Any]
    data_quality: Dict[str, Any]

For a medical world model, state representation is not a preprocessing detail.

It is the foundation.

Action: Interventions Must Become Computable Objects

In a chatbot, an intervention might appear as natural language:

Improve sleep, exercise more, and eat better.

That is not enough for a world model.

A world model needs computable action objects.

Example:

{
  "action_id": "increase_zone2_exercise",
  "type": "lifestyle",
  "target": "weekly_exercise_minutes",
  "change": {
    "from": 90,
    "to": 150
  },
  "duration": "12_weeks",
  "monitoring": [
    "resting_heart_rate",
    "sleep_quality",
    "fasting_glucose"
  ],
  "safety_notes": [
    "requires clinician review if known cardiovascular disease exists"
  ]
}

Python representation:

@dataclass
class MedicalAction:
    action_id: str
    action_type: str
    target: str
    parameters: Dict[str, Any]
    duration: str
    monitoring: List[str]
    safety_notes: List[str]

In medicine, an action could be:

a medication change;
a lifestyle intervention;
a monitoring plan;
a nutrition strategy;
a sleep intervention;
an exercise protocol;
a follow-up test;
a referral;
a decision to wait and observe.

The key is that actions should be explicit, parameterized, time-bounded, and auditable.

Evidence: Transitions Should Be Evidence-Bound

A medical world model should not freely hallucinate transitions.

Every transition hypothesis should be bound to evidence.

A minimal evidence object:

{
  "evidence_id": "evidence_001",
  "claim": "Increasing weekly aerobic exercise may improve insulin sensitivity in selected metabolic-risk populations.",
  "evidence_type": [
    "clinical_guideline",
    "peer_reviewed_study",
    "mechanistic_rationale"
  ],
  "strength": "moderate",
  "applicability": {
    "population_match": "partial",
    "condition_match": "partial",
    "uncertainty": "individual response may vary"
  },
  "limitations": [
    "not a personalized treatment prediction",
    "requires safety screening",
    "effect size depends on baseline state and adherence"
  ]
}

A simple evidence builder:

@dataclass
class EvidenceItem:
    evidence_id: str
    claim: str
    evidence_type: List[str]
    strength: str
    applicability: Dict[str, Any]
    limitations: List[str]

class EvidenceBuilder:
    def build(self, state: HealthState, action: MedicalAction):
        evidence_items = self.retrieve_relevant_evidence(state, action)
        filtered = self.filter_by_applicability(evidence_items, state)
        return filtered

    def retrieve_relevant_evidence(self, state, action):
        # In production, this should query curated knowledge bases,
        # clinical guidelines, systematic reviews, or trusted literature indexes.
        return []

    def filter_by_applicability(self, evidence_items, state):
        # Filter by population, context, condition, safety boundary,
        # measurement quality, and uncertainty.
        return evidence_items

A useful rule:

Do not generate advice. Generate evidence-bound transition hypotheses.

Transition: Hypothesis, Not Promise

In a medical world model, a transition should not be framed as a promise.

Bad framing:

This intervention will improve the outcome.

Better framing:

Given the current state and evidence constraints, this action may produce the following state changes, with uncertainty.

A transition object:

@dataclass
class TransitionHypothesis:
    from_state: HealthState
    action: MedicalAction
    expected_changes: Dict[str, Any]
    time_window: str
    evidence: List[EvidenceItem]
    uncertainty: Dict[str, Any]
    safety_flags: List[str]

Example output:

{
  "expected_changes": {
    "weekly_exercise_minutes": {
      "direction": "increase",
      "expected_from": 90,
      "expected_to": 150
    },
    "insulin_sensitivity": {
      "direction": "potential_improvement",
      "confidence": "low_to_moderate"
    },
    "fasting_glucose": {
      "direction": "possible_decrease",
      "confidence": "uncertain"
    }
  },
  "time_window": "8_to_12_weeks",
  "uncertainty": {
    "adherence": "unknown",
    "baseline_variability": "high",
    "measurement_noise": "moderate"
  },
  "safety_flags": [
    "screen cardiovascular risk before increasing exercise intensity"
  ]
}

Implementation sketch:

class TransitionModel:
    def estimate(self, state, action, evidence):
        expected_changes = self.estimate_expected_changes(
            state=state,
            action=action,
            evidence=evidence
        )

        uncertainty = self.estimate_uncertainty(
            state=state,
            action=action,
            evidence=evidence
        )

        safety_flags = self.check_safety_flags(state, action)

        return TransitionHypothesis(
            from_state=state,
            action=action,
            expected_changes=expected_changes,
            time_window=action.parameters.get("time_window", "unknown"),
            evidence=evidence,
            uncertainty=uncertainty,
            safety_flags=safety_flags
        )

The name matters.

Use TransitionHypothesis, not TreatmentPlan.

Safety Gate: Medical World Models Must Be Safety-First

A medical world model should not simulate all actions freely.

Before transition simulation, there should be a safety gate.

class SafetyGate:
    def check(self, state: HealthState, action: MedicalAction):
        checks = [
            self.check_contraindications(state, action),
            self.check_required_human_review(state, action),
            self.check_action_intensity(state, action),
            self.check_data_quality(state),
        ]

        return all(checks)

    def check_contraindications(self, state, action):
        # Placeholder only.
        # Production systems require curated medical rules and human review.
        return True

    def check_required_human_review(self, state, action):
        # Some actions should never be autonomous.
        return True

    def check_action_intensity(self, state, action):
        return True

    def check_data_quality(self, state):
        return True

Then:

def simulate_medical_transition(state, action):
    if not safety_gate.check(state, action):
        return {
            "status": "blocked",
            "reason": "Safety gate failed. Human review required."
        }

    evidence = evidence_builder.build(state, action)
    transition = transition_model.estimate(state, action, evidence)

    return transition

A useful design principle:

medical_world_model = safety_first + evidence_bound + feedback_calibrated

Feedback: Without Feedback, It Is Not a World Model

A world model should not be a one-shot answer generator.

It should support a loop:

Observe state → Define action → Simulate transition → Collect feedback → Update model

Pseudo-code:

def world_model_loop(subject, action):
    state_t0 = observe_state(subject)

    evidence = build_evidence_chain(state_t0, action)

    transition_hypothesis = simulate_transition(
        state=state_t0,
        action=action,
        evidence=evidence
    )

    feedback = collect_feedback(
        subject=subject,
        action=action,
        time_window=transition_hypothesis.time_window
    )

    updated_state = update_model(
        previous_state=state_t0,
        action=action,
        hypothesis=transition_hypothesis,
        feedback=feedback
    )

    return updated_state

Feedback can include:

repeated biomarkers;
symptoms;
wearable trends;
adherence records;
adverse events;
clinician review;
patient-reported outcomes;
environmental or lifestyle changes.

Feedback object:

{
  "feedback_id": "feedback_001",
  "action_id": "increase_zone2_exercise",
  "time_window": "12_weeks",
  "observations": {
    "weekly_exercise_minutes": 145,
    "fasting_glucose": 5.4,
    "sleep_duration": 6.5,
    "subjective_energy": "improved"
  },
  "adherence": "partial",
  "adverse_events": [],
  "notes": "Interpret carefully; multiple concurrent changes existed."
}

Update logic:

class FeedbackUpdater:
    def update(self, previous_state, action, hypothesis, feedback):
        comparison = self.compare_expected_vs_observed(
            expected=hypothesis.expected_changes,
            observed=feedback["observations"]
        )

        return {
            "previous_state": previous_state,
            "action": action,
            "hypothesis": hypothesis,
            "feedback": feedback,
            "comparison": comparison,
            "update_reason": self.infer_update_reason(comparison)
        }

Without feedback, the system cannot calibrate.

Audit Log: Every Transition Should Be Traceable

For medical AI, auditability is not optional.

A transition should have an audit trail:

{
  "audit_id": "audit_001",
  "timestamp": "2026-05-20T23:00:00+08:00",
  "state_version": "state_v1",
  "action_version": "action_v1",
  "evidence_version": "evidence_v1",
  "transition_version": "transition_v1",
  "model_version": "world_model_v0.1",
  "human_review": {
    "required": true,
    "status": "pending"
  },
  "disclaimer": "Hypothesis-generating only. Not a treatment recommendation."
}

Audit logger:

class AuditLogger:
    def log_transition(self, state, action, evidence, transition):
        return {
            "state": state,
            "action": action,
            "evidence": evidence,
            "transition": transition,
            "model_version": self.get_model_version(),
            "human_review_required": True,
            "disclaimer": "Hypothesis-generating only. Not medical advice."
        }

For medical world models, audit logs are part of the core architecture.

A Minimal Medical World Model

Putting the pieces together:

class MinimalMedicalWorldModel:
    def __init__(
        self,
        safety_gate,
        evidence_builder,
        transition_model,
        feedback_updater,
        audit_logger
    ):
        self.safety_gate = safety_gate
        self.evidence_builder = evidence_builder
        self.transition_model = transition_model
        self.feedback_updater = feedback_updater
        self.audit_logger = audit_logger

    def run(self, state, action):
        if not self.safety_gate.check(state, action):
            return {
                "status": "blocked",
                "message": "Safety gate failed. Human review required."
            }

        evidence = self.evidence_builder.build(state, action)

        transition = self.transition_model.estimate(
            state=state,
            action=action,
            evidence=evidence
        )

        audit_log = self.audit_logger.log_transition(
            state=state,
            action=action,
            evidence=evidence,
            transition=transition
        )

        return {
            "status": "hypothesis_generated",
            "transition": transition,
            "audit_log": audit_log
        }

    def update_with_feedback(self, previous_state, action, transition, feedback):
        return self.feedback_updater.update(
            previous_state=previous_state,
            action=action,
            hypothesis=transition,
            feedback=feedback
        )

Usage:

state = observe_state(subject)
action = define_action(intervention)

result = medical_world_model.run(state, action)

if result["status"] == "hypothesis_generated":
    feedback = collect_feedback(subject, action)

    updated = medical_world_model.update_with_feedback(
        previous_state=state,
        action=action,
        transition=result["transition"],
        feedback=feedback
    )

The loop is:

State → Action → Evidence → Transition → Audit → Feedback → Update

Where SteeraMed Fits

SteeraMed can be understood as a steerable biomedical world model framework.

In developer terms, it is closer to:

state-action-transition-evidence-feedback architecture

than to:

a chatbot that gives medical advice

The goal is not to automate treatment decisions.

The goal is to make biomedical AI systems more:

state-aware;
action-explicit;
evidence-bound;
feedback-calibrated;
safety-gated;
auditable;
human-reviewable.

This is especially relevant for long-term health and longevity medicine, where the problem is not a single prediction task but longitudinal state management under uncertainty.

Developer Takeaways

1. Do not start with a chatbot

A chatbot interface may be useful later, but it should not be the core architecture.

Start with state representation.

state = observe_state(subject)

2. Do not stop at risk prediction

Risk prediction is useful, but it is not a world model.

risk = predict(state)

is not enough.

3. Make actions explicit

Without explicit actions, there is no action-conditioned transition.

next_state = simulate(state, action)

4. Treat transitions as hypotheses

A medical transition is not a promise.

transition = generate_hypothesis(state, action, evidence)

5. Build evidence chains

A transition without evidence is not acceptable in medical AI.

evidence = build_evidence_chain(state, action)

6. Add safety gates before simulation

Do not simulate unsafe or unsupported actions.

if not safety_gate.check(state, action):
    require_human_review()

7. Close the loop with feedback

Without feedback, the model cannot calibrate.

model.update(feedback)

8. Log everything

Every state, action, evidence item, transition, feedback signal, and model version should be traceable.

audit_logger.log(state, action, evidence, transition, feedback)

Final Thought

A medical world model is not a larger prediction model.

It is a system architecture for reasoning about:

state + action + evidence + transition + feedback

under uncertainty.

For developers, the key shift is this:

Do not just build systems that predict what may happen.

Build systems that can represent state, encode actions, simulate evidence-bound transitions, collect feedback, and remain auditable.

That is the difference between prediction and a world model.

References and Project Links

Ha, D., & Schmidhuber, J. Recurrent World Models Facilitate Policy Evolution. Advances in Neural Information Processing Systems 31, 2018. https://papers.nips.cc/paper/7512-recurrent-world-models-facilitate-policy-evolution; arXiv version: https://arxiv.org/abs/1803.10122
LeCun, Y. A Path Towards Autonomous Machine Intelligence. OpenReview, 2022. https://openreview.net/forum?id=BZ5a1r-kVsf
Yang, Y., Wang, Z.-Y., Liu, Q., Sun, S., Wang, K., Chellappa, R., Zhou, Z., Yuille, A., Zhu, L., Zhang, Y.-D., & Chen, J. Medical World Model: Generative Simulation of Tumor Evolution for Treatment Planning. arXiv:2506.02327, 2025. https://arxiv.org/abs/2506.02327; project page: https://yijun-yang.github.io/MeWM/
Qazi, M. A., Nadeem, M., & Yaqub, M. Beyond Generative AI: World Models for Clinical Prediction, Counterfactuals, and Planning. arXiv:2511.16333, 2025. https://arxiv.org/abs/2511.16333
Katsoulakis, E., Wang, Q., Wu, H., et al. Digital twins for health: a scoping review. npj Digital Medicine, 7, 77, 2024. https://doi.org/10.1038/s41746-024-01073-0
Emmert-Streib, F., Parkkila, S., Laubenbacher, R., et al. The role of digital twins in P4 medicine: A paradigm for modern healthcare. npj Digital Medicine, 8, 735, 2025. https://doi.org/10.1038/s41746-025-02115-x
Xiong, J. World Models for Biomedicine: A Steerability Framework. Preprints.org, 2026. https://doi.org/10.20944/preprints202605.0366.v1
SteeraMed project: https://SteeraMed.com
Steerable World project: https://steerable.world

Medical AI Doesn’t Just Need Bigger Models. It Needs an ImageNet for State Transitions

JXIONG — Tue, 19 May 2026 00:15:48 +0000

Whoever builds the “state–intervention–transition” dataset for biomedicine may define the next generation of medical AI infrastructure.

Author: Jianghui Xiong

Medical AI is moving beyond classification, risk prediction, and question answering.

The next frontier is not just:


sample → label

or:


question → answer

It is:


state + action → next state

In other words:


current biological state + intervention → future biological state

To build real biomedical world models, we need more than bigger models. We need something analogous to ImageNet — not for images, but for biological state transitions.

I will call this idea, for now:


Biomedical TransitionNet

A shared infrastructure for recording, standardizing, and evaluating:


baseline biological state

- intervention
- follow-up biological state
- mechanism evidence
- uncertainty

This article explains why such an infrastructure is needed, why it matters, and why it is scientifically difficult.

It does not claim that a complete biomedical world model already exists.

1. ImageNet was not just a dataset. It was infrastructure.

When people talk about the deep learning revolution in computer vision, they often mention AlexNet, VGG, ResNet, and other neural network architectures.

That is correct, but incomplete.

One of the most important enabling factors was ImageNet.

ImageNet was not merely a large collection of images. Its deeper value was that it gave computer vision a shared coordinate system:

a common task,
a common label hierarchy,
common training and test data,
common benchmarks,
and a way to compare progress across models and institutions.

Before ImageNet, many computer vision systems were difficult to compare because they were trained and evaluated on different datasets. ImageNet helped the field converge around shared evaluation.

That is why ImageNet became much more than a database. It became research infrastructure.

Medical AI may now need something similar.

But not an image dataset.

Medicine needs an ImageNet for state transitions.

2. Medical AI has many models, but not enough transition data

Today, we already have many types of medical AI systems:

medical large language models,
medical question-answering systems,
radiology models,
pathology models,
omics foundation models,
virtual cell models,
digital twin systems,
clinical decision support tools,
AI drug discovery platforms.

These are important.

But if we think the future of medical AI is only “a bigger medical chatbot”, we may miss the real challenge.

Medicine is not only about answering questions.

Medicine is about understanding and changing biological trajectories.

A clinician does not only ask:


What disease does this patient have?

They also ask:


Why is this biological state happening?

What is driving deterioration?

Which mechanisms are actionable?

Which intervention may shift the trajectory?

How should the response be measured?

What if the expected response does not happen?

What if an adverse response appears?

These are not just language problems.

They are state transition problems.

Most medical AI today is still closer to:


sample → label

or:


question → answer

But biomedical world models require something closer to:


state + action → next state

That is the key shift.

3. What is a biomedical world model?

In AI, a world model is usually understood as an internal model that helps an agent simulate how the environment changes after an action.

A simple abstraction is:


current state + action → future state

In robotics, this may mean:


robot pose + motor command → next scene state

In autonomous driving, it may mean:


traffic scene + driving action → future traffic scene

In biomedicine, the analogous formulation would be:


biological state + intervention → future biological state

This could apply at multiple scales:


cell state + perturbation → cellular response

tissue state + treatment → tissue response

patient state + intervention → follow-up state

A biomedical world model should therefore not be understood as a medical chatbot.

It is not merely:


medical text in → medical text out

A more meaningful biomedical world model would combine:


state representation

- intervention representation
- transition modeling
- mechanism evidence
- uncertainty estimation
- feedback correction

That is much harder than ordinary medical QA.

And it requires a different kind of data.

4. Why medicine needs its own ImageNet

In computer vision, a basic supervised learning unit can often be simplified as:


image + label

For biomedical world models, the basic unit should look more like:


baseline state + action + follow-up state

Or mathematically:


S(t) + A → S(t + Δt)

Where:


S(t)       = biological state before intervention

A          = action or intervention

S(t + Δt)  = biological state after intervention

Δt         = time interval

This is fundamentally different from a static medical database.

A biomedical world model does not only need:

medical images,
electronic health records,
omics profiles,
drug-target databases,
clinical notes,
literature graphs.

Those are useful, but insufficient.

It needs structured longitudinal data describing:


what the biological state was,

what action was taken,

what changed afterward,

over what time scale,

with what evidence,

and with what uncertainty.

This is why medicine needs something like a Biomedical TransitionNet.

Not a direct copy of ImageNet.

A new infrastructure designed for biological state transitions.

5. What should one data unit look like?

A conventional supervised learning sample may look like:


x → y

Examples:


image → diagnosis label

clinical note → ICD code

genomic variant → risk category

A biomedical world-model sample should look more like:


state_before

- intervention
- state_after
- time_interval
- evidence_chain
- uncertainty

A simplified schema might look like this:


{

"baseline_state": {

"molecular": "...",

"clinical": "...",

"phenotype": "...",

"lifestyle": "...",

"context": "..."

},

"action": {

"type": "...",

"dose": "...",

"frequency": "...",

"duration": "...",

"mechanism": "..."

},

"follow_up_state": {

"molecular": "...",

"clinical": "...",

"phenotype": "...",

"adverse_events": "..."

},

"transition": {

"direction": "...",

"magnitude": "...",

"time_scale": "...",

"confidence": "..."

},

"evidence_chain": {

"target": "...",

"pathway": "...",

"biomarker": "...",

"phenotype": "...",

"validation": "..."

}

}

This is obviously simplified.

But the principle matters:

A biomedical world model should learn not only:


what this sample is

but:


how this biological system changed after a defined intervention

6. Five layers of a biomedical ImageNet

If we want to build an ImageNet-like infrastructure for biomedical world models, it should include at least five layers.

6.1 State representation

The first question is:


What is the biological state?

A patient state is not just a diagnosis label.

Terms such as:


diabetes

hypertension

aging

inflammation

fatigue

frailty

are useful, but they are high-level descriptions.

A real biological state may include:

genome,
DNA methylation,
transcriptome,
proteome,
metabolome,
immune state,
inflammatory state,
organ function,
microbiome,
sleep,
activity,
diet,
medication history,
environmental exposure,
clinical background.

A simplified representation may be:


individual_state =

molecular_state

- pathway_state
- organ_state
- phenotype_state
- lifestyle_context
- clinical_context

Without a state representation, a biomedical world model does not know what it is simulating.

6.2 Action ontology

A world model needs actions.

In medicine, actions are complex.

They may include:

drugs,
supplements,
diet,
exercise,
sleep intervention,
stress management,
cell therapy,
gene therapy,
regenerative medicine,
combination therapy,
N-of-1 personalized intervention.

Even a drug intervention requires many parameters:


drug name

dose

frequency

route

duration

combination

adherence

contraindications

adverse events

Exercise intervention also requires:


type

intensity

frequency

duration

heart-rate zone

recovery condition

baseline fitness

If actions are not standardized, the model cannot learn meaningful transitions.

6.3 Transition record

The core of a biomedical world model is the transition:


before → after

Examples:


inflammatory state before intervention → inflammatory state after intervention

DNA methylation age before intervention → DNA methylation age after intervention

metabolic state before intervention → metabolic state after intervention

tumor state before treatment → tumor state after treatment

Without follow-up measurement, there is no transition.

Without transition, there is no world model.

Many medical datasets are still one-time measurements:


one-time measurement

Biomedical world models need:


longitudinal measurement

6.4 Evidence chain

A medical model should not only output a probability.

If a model says:


This intervention may help.

That is not enough.

It should also answer:


Which targets are involved?

Which pathways are affected?

Which abnormal state does this address?

Which biomarkers can validate the response?

Which evidence comes from experiments?

Which evidence comes from clinical data?

Which part is only model inference?

Which risks should be monitored?

In medicine, prediction alone is not sufficient.

A safer output should look more like:


prediction + mechanism + validation + uncertainty

This is especially important because medical AI should not become an uninspectable black box.

6.5 Benchmark task

ImageNet helped computer vision because different models could be compared on shared tasks.

Biomedical world models need benchmarks too.

Possible benchmark tasks include:

cellular perturbation response prediction,
gene expression response after drug perturbation,
tumor state simulation after treatment,
metabolic biomarker response prediction,
inflammatory state transition prediction,
aging-related biomarker transition prediction,
N-of-1 intervention response direction prediction.

But the metrics cannot be copied directly from image classification.

Useful metrics may include:


directional accuracy

mechanistic consistency

biomarker validation

uncertainty calibration

risk awareness

cross-context generalization

This is much harder than top-1 accuracy.

But medicine requires it.

7. Related progress: promising, but still early

To be scientifically careful, we should not pretend that complete biomedical world models already exist.

They do not.

But several related directions are emerging.

7.1 ImageNet as an infrastructure analogy

ImageNet and ILSVRC showed how large-scale, standardized datasets and benchmarks can accelerate a field.

However, ImageNet is a benchmark for image classification and detection.

It is not equivalent to what biomedicine needs.

Here, ImageNet is used only as an infrastructure analogy.

The biomedical version must be longitudinal, dynamic, intervention-aware, and mechanism-sensitive.

7.2 World Models in AI

Ha and Schmidhuber’s World Models is a representative work in AI world modeling.

Its key idea is that an agent can learn an internal model of the environment and use it to simulate future states.

Medicine cannot directly copy this setting.

A human body is not a game environment.

Clinical intervention cannot be freely explored by trial and error.

But the abstraction:


state + action → future state

is still useful for thinking about medical AI.

7.3 Virtual cells and perturbation response

Arc Institute’s State model is a recent example of virtual-cell modeling.

It aims to predict how cells respond to drugs, cytokines, or genetic perturbations. Public descriptions indicate that State was trained on large-scale observational and perturbational single-cell data.

This is important because it directly touches the pattern:


cell state + perturbation → cellular response

However, State is primarily a cellular-level model.

It should not be confused with a complete patient-level biomedical world model.

7.4 Medical World Model for tumor evolution

Recent work using the term Medical World Model, such as MeWM, explores generative simulation of tumor evolution under treatment conditions.

This is relevant because it moves medical AI from static recognition toward treatment-conditioned disease dynamics.

But this direction is still early.

It should not be interpreted as a general solution to biomedical world modeling.

7.5 Digital twins and virtual physiological systems

Long before today’s AI world-model terminology, fields such as computational physiology, systems biology, virtual physiological systems, and digital twins already tried to connect biological structure, mechanism, dynamics, and measurable outputs.

That tradition matters.

A good biomedical world model should not be just a black-box predictor.

It should connect:


state

mechanism

dynamic change

measurement

feedback

Today’s biomedical world models can be seen as an extension of this older systems-modeling tradition into the era of AI, multi-omics, real-world data, and large-scale computation.

8. Why steerability matters

A biomedical world model that only predicts is not enough.

A model may predict that a patient’s risk is increasing.

But medicine needs more than that.

It needs to ask:


Which state can be measured?

Which abnormality can be explained?

Which intervention can be described?

Which transition can be tested?

Which deviation can be traced?

Which failure can be corrected?

This is why I emphasize steerability.

Going forward, I will use the name:


SteeraMed: A Steerable Biomedical World Model

Website:


https://SteeraMed.com

The earlier preprint name was:


SEWO / Steerable Medicine World Model

or in Chinese:


可驾驭医学世界模型

Whenever I mention SEWO / 可驾驭医学世界模型, it should be understood together with the new unified naming:


SteeraMed: A Steerable Biomedical World Model

The idea behind SEWO / SteeraMed is that biomedical world models should not only pursue predictive accuracy. They should also support:

state definition,
intervention description,
transition hypothesis,
mechanism audit,
deviation tracing,
uncertainty inspection,
expert steering,
and iterative correction.

The related ideas were introduced in the preprint:


World Models for Biomedicine: A Steerability Framework

and are also presented at:


https://steerable.world

Important clarification:

SEWO / SteeraMed is not a clinically validated treatment system.

It is not a medical device.

It is better understood as a structural framework and evidence-chain design principle for future biomedical world models.

The key question is not only:


Can the model predict?

but:


Can researchers and clinicians inspect, question, correct, and steer the model within clearly defined boundaries?

9. Why longevity medicine may be one entry point

Biomedical world models could start from many areas:

oncology,
cardiovascular disease,
metabolic disease,
immunology,
neurodegeneration,
drug discovery,
virtual cells,
longevity medicine.

Longevity medicine is not the only entry point.

But it is an interesting one.

Why?

9.1 Aging is a continuous state

Aging is not a single disease label.

It is a continuous, multi-system biological process involving:

inflammation,
metabolism,
immunity,
epigenetics,
mitochondrial function,
proteostasis,
stem-cell exhaustion,
cellular senescence,
organ function decline.

That makes it naturally suitable for state modeling.

9.2 Longevity medicine requires repeated measurement

Longevity medicine is not a one-time diagnostic event.

It depends on repeated measurement over time.

A useful intervention must be evaluated through:


baseline state → intervention → follow-up state

This is exactly the structure needed for biomedical world modeling.

9.3 Interventions are diverse

Longevity-related interventions may include:

diet,
exercise,
sleep,
supplements,
drugs,
cell therapy,
regenerative medicine,
stress management,
environmental exposure management.

This provides a rich action space.

9.4 Individual responses vary

The same intervention may produce different responses in different people.

That means longevity medicine cannot rely only on average effects.

It needs N-of-1 style transition data:


individual state → intervention → individual transition

Each well-structured N-of-1 intervention can be seen as a small world-model experiment.

10. Engineering implications

From an engineering perspective, the biomedical ImageNet is not just a dataset.

It is a data infrastructure problem.

It requires:

data collection,
data standardization,
multimodal integration,
time-series modeling,
intervention encoding,
causal confounding control,
privacy protection,
benchmark design,
safety boundaries,
evidence-chain tracking.

A simplified loop may look like:


measure state

↓

standardize state representation

↓

record intervention

↓

measure follow-up state

↓

construct transition sample

↓

train / evaluate world model

↓

generate testable hypothesis

↓

repeat and correct

This is not a static dataset.

It is a data flywheel.

11. Main challenges

This is scientifically and technically difficult.

Some of the main challenges include:

11.1 Biological state is complex

A human state cannot be compressed into one label.

We need ways to represent multi-omics, clinical metrics, imaging, lifestyle, symptoms, environmental exposure, and medical history as computable state variables.

11.2 Interventions are hard to standardize

Drugs, exercise, diet, sleep, supplements, and cell therapies all have complex parameters.

Without action standardization, transition learning will be noisy.

11.3 Follow-up data is scarce

Most medical data is not collected as structured pre/post intervention transition data.

This requires new data collection workflows.

11.4 Causal confounding is serious

In the real world, people often change many things at once:


diet

exercise

sleep

medication

supplements

stress

Attributing a state change to one factor is difficult.

This requires careful study design and statistical methods.

11.5 Safety and ethics are central

A biomedical world model cannot freely experiment like a game-playing agent.

Any intervention-related model must clearly distinguish:


research hypothesis

health-management suggestion

clinical decision support

medical recommendation

validated therapy

Clinical use would require prospective validation, safety evaluation, ethical review, regulatory review where applicable, and professional oversight.

11.6 Open standards and business incentives may conflict

If everything is closed, the field cannot build shared benchmarks.

If everything is open, companies may lack incentives to invest.

A practical ecosystem will need a balance among:


open benchmarks

privacy protection

commercial incentives

scientific collaboration

12. A minimal viable direction

A biomedical ImageNet should not begin by trying to simulate the entire human body.

A more realistic path is to start with minimal viable tasks.

Examples:

cellular perturbation response prediction,
tumor state change after treatment,
metabolic biomarker response prediction,
inflammatory state transition prediction,
DNA methylation age transition,
N-of-1 longevity intervention tracking.

A minimal task should define:


1. state variables
2. intervention variables
3. follow-up time
4. transition metrics
5. benchmark task
6. safety boundary

Start narrow.

Make it measurable.

Make it repeatable.

Make it auditable.

Then scale.

13. Whoever defines state, action, and transition may define the field

Medical AI will still need better models.

But bigger models alone cannot solve the problem of biomedical state transition learning.

The scarce asset is the infrastructure that allows models to learn:


how life systems change after intervention

Future platform-level medical AI companies may not be the ones with the largest language models.

They may be the ones that can build the strongest data flywheel:


measure biological state

standardize interventions

record follow-up changes

construct mechanism evidence chains

evaluate transition models

repeat

Whoever defines state defines what medical AI can see.

Whoever defines action defines how medical AI understands intervention.

Whoever defines transition defines how medical AI learns biological change.

Whoever defines the benchmark defines how the field measures progress.

Conclusion

ImageNet helped machines learn to see the world.

A biomedical ImageNet should help AI learn how life responds to intervention.

That does not mean replacing clinicians.

It means building a scientific infrastructure where models can learn:


how states form

how interventions act

how systems transition

how evidence is validated

The next decade of medical AI may not be limited by model size alone.

It may be limited by the lack of a shared infrastructure for biological state transitions.

That is the real opportunity.

References

Deng J, Dong W, Socher R, Li LJ, Li K, Fei-Fei L.

ImageNet: A Large-Scale Hierarchical Image Database. CVPR. 2009.

https://ieeexplore.ieee.org/document/5206848
Russakovsky O, Deng J, Su H, et al.

ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision. 2015.

https://arxiv.org/abs/1409.0575
ImageNet official website.

https://www.image-net.org/
Ha D, Schmidhuber J.

World Models. 2018.

https://worldmodels.github.io/
Arc Institute.

Arc Institute’s first virtual cell model: State.

https://arcinstitute.org/news/virtual-cell-model-state
Predicting cellular responses to perturbation across diverse contexts with State. bioRxiv. 2025.

https://www.biorxiv.org/content/10.1101/2025.06.26.661135v1
Yang Y, Wang ZY, Liu Q, et al.

Medical World Model: Generative Simulation of Tumor Evolution for Treatment Planning. arXiv.

https://arxiv.org/abs/2506.02327
IEEE Transactions on Biomedical Engineering.

Digital Twins / AI World Models.

https://www.embs.org/tbme/research-highlights/digital-twins-ai-world-models/
Acosta JN, Falcone GJ, Rajpurkar P, Topol EJ.

Multimodal biomedical AI. Nature Medicine. 2022.

https://www.nature.com/articles/s41591-022-01981-2
Xiong J.

World Models for Biomedicine: A Steerability Framework. Preprints.org. 2026.

https://www.preprints.org/manuscript/202605.0366

DOI: https://doi.org/10.20944/preprints202605.0366.v1
SteeraMed: A Steerable Biomedical World Model.

https://steerable.world

Disclaimer

This article is for research, technical, and industry discussion only.

It is not medical advice, diagnostic advice, or treatment advice.

Any biomedical world model intended for clinical use would require prospective validation, safety evaluation, ethical review, regulatory review where applicable, and professional clinical oversight.