The OHDSI Collaborative built something remarkable: a global network of hundreds of institutions, each holding patient data in a common format (OMOP CDM), each capable of running the same analysis and returning comparable results — without ever transferring a single patient record.
That is a genuinely hard coordination problem. The vocabulary standardization alone took years. SNOMED CT, RxNorm, LOINC, ICD-10 — all mapped, harmonized, and maintained across DataPartners on five continents. ATLAS translates research questions into executable queries. HADES packages and distributes them. ACHILLES characterizes each node. This infrastructure represents an enormous investment by the scientific community, and it works.
But there is a gap between what the OHDSI network does and what it could do — and that gap is not a criticism of OHDSI's design. It is a structural feature of how federated query networks accumulate knowledge.
Christopher Thomas Trevethan discovered the Quadratic Intelligence Swarm (QIS) protocol and filed 39 provisional patents on the architecture. The more I have studied how OHDSI works, the more I believe the OHDSI network is one of the most natural environments for QIS outcome routing to demonstrate its value. This article is an open invitation to OHDSI researchers to evaluate whether that is true.
The Gap: Studies Teach the Coordinating Center, Not Each Other
Here is how knowledge currently flows in a typical OHDSI distributed study:
- A coordinating center identifies a research question
- ATLAS generates a cohort definition
- DataPartners run the analysis locally
- Aggregate results return to the coordinating center
- The coordinating center synthesizes and publishes
The intelligence flows from nodes to center. The center publishes a paper. Future researchers can read that paper. But the DataPartner in Des Moines that contributed to the 2024 ACE inhibitor study cannot automatically incorporate what the DataPartner in Dublin learned about the same drug class in their 2025 renal outcomes study. Each study cycle starts fresh. The network learns; the nodes do not.
This is not a flaw in OHDSI's design. OHDSI was architected to enable reproducible, privacy-preserving distributed research — and it does that. But the accumulation of intelligence across DataPartners, in real time, between studies, is not what the current architecture was built for.
That is exactly what QIS outcome routing does.
What QIS Adds to OMOP CDM Networks
QIS distributed outcome routing is not a replacement for OHDSI's query infrastructure. It is an additional layer that runs between the DataPartner's local analysis and the coordinating center's synthesis — adding continuous cross-node learning to the existing batch query model.
The architecture maps cleanly to OHDSI's existing components:
OMOP CDM vocabularies → QIS semantic fingerprints
Each OMOP CDM DataPartner already encodes patient events using controlled vocabularies: SNOMED concept IDs for conditions, RxNorm concept IDs for drug exposures, LOINC concept IDs for measurements. In a QIS outcome routing layer, these concept IDs compose directly into a semantic fingerprint — a vector representation of the clinical context that produced a given outcome.
The fingerprinting work is already done. Every OMOP CDM DataPartner already speaks the same semantic language. QIS can use that shared vocabulary as its addressing system without requiring any new ontology work.
ACHILLES characterization → QIS node profile
ACHILLES generates data characterization reports for each DataPartner: population demographics, condition prevalence, measurement distributions. This characterization data is exactly the kind of information a QIS node uses to route its outcome packets to semantically similar nodes.
An institution specializing in rare renal conditions has a distinctive ACHILLES fingerprint. So does a pediatric specialty center. So does a community hospital serving a primarily elderly population. In a QIS routing layer, DataPartners with similar ACHILLES profiles automatically become each other's closest semantic neighbors — which means their outcome packets route to each other first.
Distributed phenotyping → QIS continuous outcome emission
OHDSI's distributed phenotyping validation framework (PHOEBE, PheValuator) generates phenotype algorithm performance metrics across DataPartners: sensitivity, specificity, positive predictive value. These are exactly the kind of outcome data QIS is designed to route.
A DataPartner that has validated a phenotyping algorithm for Type 2 Diabetes against their local gold standard can emit an outcome packet containing: algorithm ID, performance metrics, population context, validation approach. That packet routes to other DataPartners attempting phenotype validation for similar conditions — before the coordinating center's next query cycle even begins.
The Quadratic Scaling Argument for OHDSI
The OHDSI network currently has more than 380 DataPartners. The math of QIS distributed outcome routing at OHDSI scale is straightforward:
N(N-1)/2 synthesis opportunities
- 380 DataPartners = 72,010 unique pairwise synthesis opportunities
- Every pairwise synthesis opportunity is: one DataPartner's validated outcome informing another DataPartner with a similar clinical context
Under the current batch query model, those 72,010 pairwise learning opportunities are only partially realized — and only when the coordinating center runs a study that happens to include both DataPartners. Under QIS continuous routing, they can activate in real time, as outcome packets flow between semantically similar nodes.
The compute cost does not scale at N². Each DataPartner only processes the packets relevant to its clinical context. The routing mechanism — whether implemented as a DHT, a database semantic index, or a vector similarity search — handles addressing at O(log N) or better. The network gets quadratically smarter while each node pays a logarithmic compute cost.
The Rare Disease Problem
One of the most significant practical limitations of the current OHDSI model is the statistical disclosure threshold: DataPartners with fewer than a minimum cell count (typically N=5 to N=11, depending on the DataPartner's policy) must suppress their results. For rare disease studies, this means that the institutions with the deepest expertise — rare disease specialty centers — may be among the least able to contribute to federated OHDSI queries.
QIS outcome routing handles this differently. The constraint is not on data counts; it is on outcome packets. A rare disease specialty center with 3 patients who match a complex phenotype can emit an outcome packet describing what worked for those 3 patients — the clinical approach, the measurement context, the outcome trajectory — without disclosing cell counts or patient-level records. That packet routes to the 2 other specialty centers in the world with similar patients.
The federated learning analogy is exact: federated learning also struggles with N=1 and N=2 sites because there is not enough local data to compute a meaningful gradient. QIS does not require gradient computation. It requires only that an outcome happened and can be described in ~512 bytes.
For the OHDSI researcher working on a condition with 200 known cases in the entire OHDSI network, this matters.
The Technical Reference
For OHDSI researchers who want the architecture detail — how OMOP concept IDs compose into semantic fingerprints, what a QIS outcome packet looks like in OMOP context, how the routing layer integrates with existing ATLAS/HADES infrastructure — there is a dedicated technical reference:
QIS Protocol: A Technical Reference for OMOP CDM and OHDSI Network Routing
That article covers the OHDSI architecture in detail, maps QIS components to OMOP CDM components, and includes a worked example of how a phenotyping outcome packet routes through a QIS layer to OHDSI-adjacent nodes.
An Open Invitation
I know from our reader analytics that OHDSI researchers are reading QIS content — from OHDSI annual conference cities, from institutions active in the OHDSI community, from researchers running OMOP-adjacent searches that surface our architecture articles.
If you are an OHDSI researcher and you have read this far: I am interested in your evaluation. Not a pitch — a genuine evaluation.
The questions I would want you to ask:
Does the OMOP vocabulary → semantic fingerprint mapping hold? Is there information loss in composing OMOP concept IDs into a routing vector that would limit QIS's utility in a real DataPartner context?
Is the rare disease participation argument actually novel? Are there existing OHDSI mechanisms I am not aware of that already solve the N=1 site participation problem without QIS outcome routing?
Where would the coordinating center's role change? QIS distributed outcome routing does not eliminate the coordinating center — it changes what the coordinating center synthesizes. Is that a design constraint or a feature for your network?
The QIS protocol was discovered by Christopher Thomas Trevethan and carries 39 provisional patents covering the architecture. The humanitarian licensing structure makes it free for academic research and education — which is exactly where OHDSI operates. Commercial licensing revenue funds deployment to underserved health systems globally.
If you are evaluating QIS for a distributed phenotyping, real-world evidence, or cross-institutional learning context: the technical reference above is the right starting point. If you have questions or want to discuss the architecture for your specific OHDSI use case, the qisprotocol.com contact page reaches Christopher's team directly.
The OHDSI network already has the semantic standardization. It already has the DataPartner infrastructure. It already has the commitment to privacy-preserving distributed analysis. QIS outcome routing is the continuous learning layer that the current architecture is missing.
That is not a criticism. It is an invitation.
Christopher Thomas Trevethan discovered the Quadratic Intelligence Swarm (QIS) protocol on June 16, 2025. QIS scales intelligence quadratically — N(N-1)/2 synthesis opportunities — while compute scales logarithmically. 39 provisional patents filed. Free for humanitarian, research, and educational use.
Previous in this series: QIS Protocol: A Technical Reference for OMOP CDM and OHDSI Network Routing
Top comments (0)