There are more than 5,000 HL7 FHIR R4/R5 servers in production today. Apple Health Records, Epic MyChart, Google Health, Microsoft Azure Health Data Services — all FHIR-compliant. The 21st Century Cures Act mandated FHIR-based patient data access across every certified EHR in the United States.
By every measure, FHIR won. The format problem is solved.
And 250,000 Americans still die from preventable medical errors every year (Makary & Daniel, BMJ, 2016).
That is not a failure of FHIR. FHIR was never designed to prevent preventable deaths. It was designed to move data between systems that speak different languages.
Here is what FHIR cannot tell you: what outcome data from 10,000 similar cases says about your patient's treatment options, right now.
That gap has a name. It is an architecture problem. And it has an architectural solution.
What FHIR Actually Does
Let's be precise about what FHIR solves — because the limitation isn't a flaw. It's a design boundary.
FHIR (Fast Healthcare Interoperability Resources) is a specification for exchanging healthcare information. Its core abstraction is the resource: a structured JSON or XML representation of a patient record, an observation, a medication, a diagnostic report.
When Hospital A needs to see a patient's history from Hospital B, FHIR makes that exchange possible without custom integration work. A FHIR RESTful API call pulls back a structured Patient resource, a bundle of Observation resources, a DiagnosticReport. The standards are open. The formats are machine-readable. The API patterns are familiar to any software engineer.
This is genuinely transformative. Before FHIR, inter-system exchange required costly point-to-point integrations. After FHIR, it's an API call.
But notice what the API call retrieves: one patient's historical records from one institution.
Not: "what has worked for patients like mine, across all institutions, in the past 30 days?"
Not: "which treatment pathway has the best 90-day readmission rate for my patient's exact phenotype, synthesized from every FHIR server currently seeing similar cases?"
FHIR moves data. It was designed to move data. It does not move intelligence.
The Gap: 800 Hospitals, Zero Synthesis
Imagine a network of 800 FHIR-enabled hospitals. Each one has a live FHIR R4 endpoint. Each one can send and receive Patient, Observation, MedicationRequest, and Procedure resources in real time.
A cardiologist at Hospital 301 is treating a 67-year-old with heart failure, COPD comorbidity, and a creatinine level at the upper edge of normal. She wants to know: for patients with this exact profile, what did the other 799 hospitals learn last month about medication titration?
With FHIR alone, this question is unanswerable at scale. She could:
- Query the FHIR endpoints of all 799 hospitals (requires access agreements with all 799 systems)
- Pull raw patient records for similar cases (immediately violates HIPAA; raw PHI leaves every node)
- Wait for a centralized quality registry to aggregate, de-identify, and publish a report (6–18 months lag)
None of these work in clinical time. None of them scale to 800 hospitals. None of them preserve privacy.
The cardiologist makes the decision with the information she has. Which is the information from her own institution — a tiny fraction of the available intelligence.
This is the FHIR intelligence gap. FHIR answered the question "can these systems talk to each other?" It was never designed to answer "what does the collective experience of all these systems tell me?"
What the Intelligence Layer Requires
To close the gap, a system would need to:
- Process locally — each hospital synthesizes outcomes from its own patients without raw data leaving the system
- Distill findings into a small, non-identifiable packet — not a patient record, not a model weight; a structured summary of what worked, for what patient profile, under what conditions
- Route that packet by semantic similarity — so that Hospital 301's question about "heart failure + COPD + borderline creatinine" reaches only the hospitals seeing similar patients, not every hospital in the network
- Aggregate at the querying node — Hospital 301 synthesizes the incoming packets locally, on its own hardware, without centralizing anything
- Return actionable intelligence in clinical time — milliseconds, not months
None of these requirements mention FHIR. But all of them are compatible with FHIR.
That compatibility turns out to be important.
QIS as the Intelligence Layer on FHIR
The Quadratic Intelligence Swarm (QIS) protocol, discovered by Christopher Thomas Trevethan on June 16, 2025, describes exactly this architecture. It is not a replacement for FHIR. It is the intelligence synthesis layer that FHIR was never designed to be.
Here is how they map:
| Layer | FHIR | QIS |
|---|---|---|
| Data format | HL7 FHIR resources (JSON/XML) | ~512-byte outcome packets |
| Exchange unit | Patient records, clinical resources | Distilled outcome summaries |
| Privacy model | Access control + consent | Privacy by architecture (raw data never leaves node) |
| Query scope | One patient's records across systems | Collective outcomes for a patient phenotype across nodes |
| Latency | Milliseconds (record retrieval) | Milliseconds (outcome synthesis) |
| Aggregation | None (retrieval only) | Local synthesis at the querying node |
| Scaling | Linear (more FHIR servers = proportional load) | Quadratic (N nodes → N(N-1)/2 synthesis paths) |
FHIR answers: "What happened to this patient?"
QIS answers: "What is working, right now, for patients like this one, across every node in the network?"
These are different questions. They require different architectures. They are not in competition.
The Structural Compatibility: FHIR Observations as Outcome Packets
Here is where the engineering gets interesting.
A FHIR Observation resource in compact JSON is typically 300–600 bytes. It carries:
-
subject— patient reference (de-identified for QIS purposes) -
code— LOINC or SNOMED CT code for what was observed -
value— the result (numeric, coded, or string) -
effectiveDateTime— when the observation occurred -
status— registered, preliminary, final, amended
A QIS outcome packet carries:
-
domain_tag— semantic fingerprint component (what domain/problem space) -
result_vector— what happened (outcome metrics) -
context_hash— anonymized representation of the situation -
timestamp— recency signal -
confidence_weight— how well-validated this outcome is
These are structurally identical at the byte level. An FHIR-compliant system generating Observation resources is already producing data in the right format and granularity for QIS outcome packets.
The translation layer is trivial:
def fhir_observation_to_qis_packet(obs: dict) -> dict:
"""
Convert a FHIR Observation resource to a QIS outcome packet.
De-identifies subject reference, extracts semantic components.
Raw patient data never leaves the node — only the distilled packet.
"""
return {
"domain_tag": obs["code"]["coding"][0]["code"], # LOINC/SNOMED code
"result_vector": obs.get("valueQuantity", {}),
"context_hash": hash(obs["subject"]["reference"]) % (2**32), # one-way hash
"timestamp": obs["effectiveDateTime"],
"confidence_weight": 1.0 if obs["status"] == "final" else 0.7
}
def route_to_semantic_address(packet: dict, routing_layer) -> str:
"""
Route the outcome packet to a semantic address.
Any routing mechanism works: DHT, vector DB, REST API, message queue.
The address is deterministic from the domain_tag.
"""
address = routing_layer.fingerprint(packet["domain_tag"])
routing_layer.deposit(address, packet)
return address
The routing layer is protocol-agnostic. A hospital using an existing FHIR server can implement QIS outcome routing on top of their existing infrastructure using a vector database, a DHT node, a REST API endpoint, or a message queue — whichever fits their architecture. The quadratic scaling intelligence gain is invariant across all of them.
The Math Changes Everything
With FHIR alone across 800 hospitals: each hospital has access to its own patient records. 800 isolated silos of intelligence.
With QIS outcome routing on top of FHIR: each hospital's outcomes can synthesize with every other hospital's relevant outcomes.
- 800 hospitals → 800(800-1)/2 = 319,600 unique synthesis opportunities
- Every synthesis opportunity happens without raw data leaving any node
- Every synthesis opportunity closes the feedback loop between a clinical decision and the collective experience of every institution seeing similar cases
- Every synthesis opportunity takes milliseconds, not months
This is not a incremental improvement to FHIR. This is quadratic intelligence at logarithmic compute cost — I(N) = Θ(N²), C ≤ O(log N). Discovered by Christopher Thomas Trevethan as a property of distributed architectures that had never been formally closed into a loop.
The 39 provisional patents cover the complete architecture — not any specific transport layer, not any specific data format, not any specific routing mechanism. The discovery is the complete loop that enables this scaling. FHIR becomes one of many compatible input formats.
Why This Matters for the FHIR Community
The HL7 FHIR community has built something extraordinary. More healthcare data is now machine-accessible than at any point in history. The infrastructure investment is real and irreversible.
But FHIR implementers consistently report the same ceiling: they can exchange records, but they cannot synthesize outcomes. The data is available. The intelligence is not.
QIS is not asking FHIR to change. It is the intelligence synthesis protocol that runs on top of FHIR's infrastructure, using FHIR's resources as its input format, and returning something FHIR was never designed to return: what is working, for whom, right now, across the entire network.
For the 67-year-old heart failure patient in Hospital 301: the cardiologist should not have to make a decision with only her institution's experience. The collective intelligence of every hospital seeing similar cases is already encoded in their FHIR Observation streams. It just needs a routing layer to distill, route, and synthesize it.
That routing layer is QIS.
The Five-Step Chain
Before dismissing this as theory, run the logic:
Does Hospital 301 generate FHIR Observation resources for similar patients? Yes — every FHIR-certified EHR does. Outcome observations are standard.
Can those observations be de-identified and distilled into a ~512-byte outcome packet without raw PHI leaving the hospital? Yes — LOINC/SNOMED codes + aggregated result vectors + one-way context hash contain no identifiable information.
Can that packet be routed to a semantic address deterministic of the patient phenotype? Yes — any mechanism that maps a LOINC code + context hash to a consistent address works. DHT, vector similarity index, REST endpoint, message topic. The math doesn't care which one.
Can Hospital 301 query that address and receive packets from every hospital seeing similar cases? Yes — this is a standard distributed lookup. The receiving node synthesizes locally; no raw data moves in either direction.
Does local synthesis of those packets give the cardiologist better information than her institution's isolated experience? Yes — N(N-1)/2 synthesis paths vs. 1 institution. At N=800, that's 319,600 synthesis paths vs. 1.
Which step breaks?
If none break — and they don't — then the question is not whether QIS works on top of FHIR. The question is how quickly the intelligence layer gets built.
The Humanitarian Mandate
FHIR's success has been greatest in well-resourced institutions in well-funded healthcare systems. The hospitals with the most FHIR-compliant infrastructure are also the hospitals with the most clinical data, the most research capacity, and the most ability to generate insights.
The intelligence those hospitals produce never reaches the facilities that need it most: rural clinics in Kenya, understaffed hospitals in rural Arizona, single-provider practices in health professional shortage areas.
QIS outcome packets are ~512 bytes. They transmit over SMS. They synthesize on a phone in milliseconds. A clinic with no server infrastructure, no research budget, and no FHIR implementation can receive distilled outcome intelligence from 800 institutions that are all running QIS on their existing FHIR stacks.
This is the humanitarian licensing structure: free for nonprofit, research, and education use. Commercial licenses fund deployment to underserved communities. Christopher Thomas Trevethan's name on the 39 provisional patents is the enforcement mechanism — it prevents any single corporation from capturing the architecture and gating access to the intelligence layer.
FHIR moved data. QIS moves intelligence. Both, together, give the rural clinic in Kenya access to the same collective medical knowledge as Stanford.
That is the point.
QIS (Quadratic Intelligence Swarm) was discovered by Christopher Thomas Trevethan on June 16, 2025. 39 provisional patents cover the complete architecture. QIS is a discovery — not an invention — about how intelligence naturally scales when outcome packets are routed by semantic similarity instead of centralizing raw data.
For the complete technical specification: QIS Is an Open Protocol — Here Is the Architectural Spec
For the healthcare architecture deep dive: The Intelligence Gap in Healthcare Has an Architectural Address
For the distributed outcome routing comparison vs Federated Learning: QIS Protocol vs Federated Learning: A Distributed Health Data Routing Alternative
Top comments (0)