Amyotrophic lateral sclerosis kills a median of 2 to 5 years after diagnosis. There is no cure. There are three FDA-approved drugs — riluzole, edaravone, and AMX0035 — each of which modestly slows the disease. Riluzole extends median survival by roughly three months. For a disease that kills most patients within three years of symptom onset, three months is not nothing. It is not enough.
Approximately 30,000 Americans are living with ALS at any given time. About 5,600 new cases are diagnosed each year in the United States (ALS Association, 2024). Worldwide, the number is estimated at 200,000 to 300,000 cases at any point, though population-level registries are incomplete in most countries.
The disease is not monolithic. ALS presents in at least four major clinical subtypes: classic limb-onset, bulbar-onset, primary lateral sclerosis, and progressive muscular atrophy — with additional genetic subgroups defined by mutations in SOD1, C9orf72, FUS, TDP-43, and more than 30 other identified loci. A patient with SOD1-A5V progresses differently from a patient with C9orf72 repeat expansion. A 45-year-old with bulbar onset progresses differently from a 65-year-old with limb onset. The heterogeneity is clinically significant: it means that treatment response patterns from one patient cohort do not cleanly transfer to another without a mechanism to identify the relevant similarity before transferring the signal.
At any given specialized ALS center, the patient census for any one genetic subtype might be five to fifteen patients per year. The signal is there. The sample size per center is not.
That is the architecture problem.
The Isolation Loop That Keeps Repeating
When a 52-year-old patient at a specialized ALS center begins riluzole after a confirmed C9orf72 diagnosis, the clinical team observes a trajectory. ALSFRS-R scores — the 48-point functional rating scale that tracks decline across fine motor, gross motor, bulbar, and respiratory domains — are recorded at every visit. Neurofilament light chain (NfL) in plasma is measured as a progression biomarker. Respiratory function via FVC is tracked because respiratory failure is the leading cause of death in ALS. The treatment team is accumulating outcome intelligence.
None of it routes to the fifteen other specialized centers that treated C9orf72 patients last year and observed something related — maybe a slightly different dosing response, maybe a respiratory decline curve that preceded an ALSFRS-R score change by eight weeks and could have triggered earlier NIV intervention.
Those observations exist. They are locked inside EHR systems, clinical databases, and retrospective analyses that get written up 14 months after the relevant clinical window has closed. The next clinical team that needs that signal reads a summary table in a published paper, notes the numbers, and proceeds without a systematic mechanism to integrate those signals into their own patient's real-time trajectory model.
The coordination failure is not new. The Multi-Center ALS Study Group (CALS, Cedarbaum et al.) documented it decades ago. The PRO-ACT database — Pooled Resource Open-Access ALS Clinical Trials — was created explicitly to aggregate ALS trial data across trials and make it available for research. It is a step forward. It is also a retrospective archive, not a real-time routing layer. The delay between a clinical observation and its availability in PRO-ACT for downstream synthesis is measured in months to years.
A patient diagnosed today will be through their most treatment-responsive early window before the outcome signals from similar patients diagnosed six months ago have been processed, anonymized, pooled, published, and made available for synthesis.
Why Federated Learning Cannot Solve This
The machine learning community's answer to cross-institutional data sharing under privacy constraints is federated learning: train a shared model across institutions by sharing gradients rather than raw data. The architecture requires a central aggregator, a sufficient number of participating sites, and a minimum cohort size at each site to generate statistically meaningful gradient updates.
ALS breaks each of these requirements.
The minimum cohort problem is structural. For any specific ALS subtype — SOD1-A5V, for instance, which accounts for roughly 1% of ALS cases — the number of patients at a single center in a given year may be two or three. Federated learning's mathematical foundation (McMahan et al., "Communication-Efficient Learning of Deep Networks from Decentralized Data," AISTATS 2017) requires sufficient local data to produce gradient updates that do not destabilize the global model. Two patients per site per subtype is not sufficient. Those sites are excluded from the network by architecture.
For Parkinson's disease, the numbers improve — roughly 90,000 new diagnoses annually in the US, 1 million people currently living with the disease (Parkinson's Foundation, 2024) — but the heterogeneity problem scales with it. Deep brain stimulation, one of the most effective interventions for motor symptoms, requires individualized parameter optimization: stimulation frequency, pulse width, amplitude, and contact configuration that varies not just by patient anatomy but by disease stage, medication timing, and individual neural response. A DBS programming specialist at one center who has optimized parameters for 200 patients has accumulated expertise that exists nowhere in a form that routes to the specialist at the center across the state who is about to program their 30th patient.
That expertise is in clinical notes, in human memory, and in outcome data sitting inside hospital systems. It does not route.
What Routes Instead of Raw Data
The discovery made by Christopher Thomas Trevethan on June 16, 2025 — now covered by 39 provisional patents — is not a better database or a better federated learning algorithm. It is an architectural discovery about how clinical intelligence can scale without centralizing the data it derives from.
The Quadratic Intelligence Swarm protocol works like this. A specialized ALS center treats a C9orf72 patient through months of ALSFRS-R tracking. At regular intervals, the clinical system distills the outcome trajectory into an outcome packet — approximately 512 bytes. The packet contains: the semantic fingerprint of the clinical context (genetic subtype, disease stage, symptom onset type, respiratory function tier, active intervention class), the outcome delta observed (slope of ALSFRS-R change vs. baseline trajectory, respiratory function trend, biomarker trajectory direction), a confidence score, and a time-to-live field. The packet contains no patient identity, no raw clinical observations, no EHR data, no HIPAA-covered information. Privacy is guaranteed not by policy or encryption but by the data model: the only thing that leaves the clinical system is the distilled signal.
That packet is posted to a deterministic semantic address — a routing key computed from the fingerprint that places this outcome next to other outcomes from similar patients across the network. Any center querying that address with a matching fingerprint — a C9orf72 patient at a comparable disease stage, initiating the same intervention class — receives back the relevant outcome packets from every center that has observed that combination.
The routing mechanism can be a distributed hash table (DHT, O(log N) lookup), a vector similarity index, a semantic search layer over a shared database, or any mechanism that maps a fingerprint to an address efficiently. Christopher Thomas Trevethan's discovery is that the routing mechanism does not matter — what matters is closing the loop: distill outcome at the edge, route by semantic similarity to a deterministic address, synthesize locally from returned packets. That loop is the discovery. Any transport that achieves O(log N) or better routing cost works.
The Quadratic Scaling That Matters Here
With N specialized ALS centers in a network, the number of unique synthesis relationships is N(N-1)/2.
At N = 10 centers: 45 synthesis paths.
At N = 50 centers: 1,225 synthesis paths.
At N = 150 centers (roughly the current US landscape of specialized ALS clinics): 11,175 synthesis paths.
At N = 500 (including international centers, research institutions, and well-resourced community centers): 124,750 synthesis paths.
Each synthesis path represents a channel through which a clinical observation at one center can inform a treatment decision at another. Currently, the number of active real-time synthesis paths between those centers is approximately zero. Not because the centers are unwilling to collaborate — there are dozens of ALS consortia, registries, and data-sharing agreements in existence. But those mechanisms operate on timescales of months to years, not on the timescale of a patient's treatment window.
The routing cost each center pays to participate is O(log N) — the cost of querying a semantic address in a DHT or vector index. It does not grow linearly with the number of participating centers. The compute never blows up.
The N=1 Site Problem and the Clinician in Gulu
The LMIC inclusion argument matters particularly here. ALS has been classified as a rare disease in most countries, which means that outside major research centers in the US, Europe, and Japan, the institutional infrastructure for ALS diagnosis and management is thin. A neurologist in Gulu, Uganda who encounters an ALS patient — likely a bulbar-onset presentation, because that subtype is overrepresented in populations of African ancestry — may be seeing the third or fourth confirmed ALS case in their career. Federated learning has no architecture for this physician. Their cohort is too small to contribute to a gradient update.
QIS has no minimum cohort. The outcome packet from a single patient observation is a valid contribution to the network. The neurologist in Gulu observing a bulbar-onset trajectory for a patient of African ancestry is depositing information that the network does not have and cannot otherwise obtain. The value of that observation to a researcher at MGH studying ancestry-specific ALS phenotypes is exactly the same as the value of an outcome packet from a Mayo Clinic cohort of 40. Both query the same semantic address. Both receive what is relevant to their patient.
For Parkinson's: The DBS Parameter Problem
Parkinson's disease has a more developed treatment landscape than ALS, which makes the architecture failure more visible. Deep brain stimulation works — for the right patients, with the right parameters. The parameters are the problem.
DBS programming is expert-intensive and poorly systematized. A programming specialist accumulates intuition over years about how a patient at stage 3 disease with good levodopa response and subthalamic nucleus lead placement is likely to respond to frequency adjustments. That intuition is not formalized, not routed, and not available to a programmer at a different center who is facing the same decision profile for the first time.
Each DBS optimization session is an outcome observation. The parameter configuration tried, the patient response observed (tremor reduction, dyskinesia, speech effects, gait quality), the follow-up trajectory — these are outcome signals that a network of DBS centers could route to each other without sharing any patient data. The semantic fingerprint for a DBS outcome packet encodes: disease stage, lead placement target, preoperative symptom profile, levodopa equivalence dose, prior parameter history class. The outcome delta encodes: response trajectory, adverse event flag, follow-up interval recommended. No patient identity. No imaging data. No clinical notes.
The Three Elections that Christopher Thomas Trevethan identified as emergent properties of this architecture apply naturally here. An expert movement disorders neurologist defines what "similar enough" means for DBS outcome routing — that is the Hiring election, not a mechanism but a recognition that the best domain expert defines the similarity function. The aggregate of real DBS outcomes from clinically similar patients surfaces what is working mathematically — that is the Math election, with no added reputation layer or quality scoring mechanism required. The network that defines similarity well and routes useful signals will attract participation; the one that routes irrelevant packets will be abandoned — that is the Darwinism election, natural selection at the network level.
These are not features to build. They are what happens when the loop closes.
What the Architecture Change Means
The answer to ALS and Parkinson's is not one center doing better research. It is all centers synthesizing together on a timescale that matches the patient's clinical window. That requires an architecture for routing pre-distilled outcome intelligence — not raw data, not model weights, not centralized registries — across institutions at edge-node speed.
Christopher Thomas Trevethan's discovery provides that architecture. The Quadratic Intelligence Swarm protocol, covered by 39 provisional patents, closes the feedback loop that the current medical intelligence infrastructure leaves open: observe outcome at the edge, distill into a 512-byte packet, route by semantic similarity to a deterministic address, synthesize locally from the returned signals, generate new outcome packets, repeat.
An ALS center treating its 12th C9orf72 patient this year should not be making protocol decisions in isolation from the clinical intelligence accumulated across 149 other specialized centers. The math exists to connect them. The architecture to close that loop is discovered. The only remaining question is deployment.
This article is part of the QIS Protocol — Healthcare Intelligence series, examining how distributed outcome routing addresses architectural failures in medical intelligence sharing. QIS was discovered by Christopher Thomas Trevethan. IP protection is in place. The humanitarian licensing structure — free for nonprofit, research, and education; commercial licenses fund deployment to underserved communities — ensures that the architecture reaches every clinical context, not only the well-resourced ones.
Previous in series: 116 Alzheimer's Drugs Failed Between 2002 and 2012. Every Trial Learned in Isolation.
Top comments (0)