DEV Community

JXIONG
JXIONG

Posted on

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

#ai #worldmodel #steeramed #steerability

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:

  1. Why report-reading medical AI is not enough.
  2. What a medical world model is.
  3. How State-Action-Transition-Evidence-Feedback can describe biomedical reasoning.
  4. How SteeraMed Core works as an early medical world model prototype.
  5. How the SEMO algorithm supports candidate intervention reasoning.
  6. 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?
Enter fullscreen mode Exit fullscreen mode

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?
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

In plain language:

current state + current action -> next state
Enter fullscreen mode Exit fullscreen mode

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

patient_state_t + intervention_action_t -> patient_state_t+1
Enter fullscreen mode Exit fullscreen mode

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?
Enter fullscreen mode Exit fullscreen mode

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?
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

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?
Enter fullscreen mode Exit fullscreen mode

or:

Is this person high-risk, medium-risk, or low-risk?
Enter fullscreen mode Exit fullscreen mode

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?
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

SEMO emphasizes a different chain:

individual omics state
-> network deviation modules
-> candidate action networks
-> mechanism evidence
-> feedback metrics
Enter fullscreen mode Exit fullscreen mode

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
Enter fullscreen mode Exit fullscreen mode

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?
Enter fullscreen mode Exit fullscreen mode

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.

Related links:

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

Top comments (0)