The theme of this week's European OHDSI Symposium in Rotterdam is Continuous Collaboration for Living Evidence Generation.
It is exactly the right theme. And it points directly at the one structural gap the OHDSI network has not yet closed.
OHDSI has done something remarkable: 400+ sites across dozens of countries, all running the OMOP Common Data Model, all capable of participating in the same federated study, all speaking the same clinical data language. The network works. The evidence it generates is real and increasingly influential in regulatory and clinical settings.
But the phrase living evidence is aspirational, not yet architectural. Here is why — and here is what closes the gap.
What "Living Evidence" Requires
Living evidence is not the same as periodic federated studies. It means:
Evidence that updates as clinical reality changes. Not monthly. Not per-study-cycle. Continuously — as outcomes are generated by the sites producing them.
Evidence that synthesises across sites without requiring per-synthesis coordination. A new treatment combination emerging at five European sites simultaneously should propagate intelligence to all 400 sites before the next planned study is designed, not after it completes.
Evidence accessible to sites with small populations. A site with 12 patients presenting a rare phenotype cannot meaningfully contribute to a gradient-based federated analysis. It can contribute one outcome packet.
Current OHDSI architecture enables the first kind of evidence generation — planned, coordinated, federated — with exceptional rigour. It does not yet have the architectural layer for the second and third kinds.
The 79,800 Number
The OHDSI network has approximately 400 participating data sites as of 2026. Some analyses cite more, some fewer depending on the study, but 400 is a reasonable working figure for the European + global network scope.
At 400 sites: N(N-1)/2 = 79,800 potential synthesis pathways.
A synthesis pathway is what happens when Site A's distilled outcome from treating patients with condition X informs Site B's treatment decisions for patients with the same condition X profile — in near real time, without a study being launched.
Under current OHDSI architecture, that pathway count is not 79,800. It is the number of active federated studies running at any given moment — typically a handful — and each pathway requires months of coordination, protocol design, IRB approvals, and post-hoc analysis before any site receives any synthesis.
Not because OHDSI is doing something wrong. Because no architectural layer exists yet to route pre-distilled outcomes between sites continuously. That is not a criticism of OHDSI. It is a description of a missing infrastructure layer.
The Discovery Behind the Architecture
On June 16, 2025, Christopher Thomas Trevethan discovered an architecture that makes continuous outcome routing possible.
The discovery is called Quadratic Intelligence Swarm (QIS) Protocol. It is filed under 39 provisional patents. The breakthrough is not a component — it is the complete architecture loop:
Raw clinical signal → local processing → distilled outcome packet (~512 bytes) → semantic fingerprinting → routing to a deterministic address defined by the clinical problem class → delivery to all sites sharing that problem class → local synthesis → new outcome packets → loop continues.
No raw patient data moves. No model weights are shared. No central aggregator is required. No per-study coordination is needed for the routing layer — though planned studies remain valuable and complementary.
Each site processes its local OMOP data and emits an outcome packet: treatment pathway, outcome class, patient phenotype profile, confidence interval. The packet is semantically fingerprinted and routed to the address every site with the same clinical problem class is listening to.
Every site with that problem class receives every relevant outcome packet from every other site, automatically, in near real time. They synthesise locally. Their own analysis improves without any patient data leaving.
Why This Is Not Federated Learning
The comparison to federated learning is worth making precisely, because it matters for OHDSI's architecture decisions.
Federated learning:
- Shares model gradients (not patient outcomes)
- Requires a central aggregator to collect and average gradients
- Requires sufficient local data at each site for meaningful gradient computation — which excludes N=1 and N=2 sites
- Operates in rounds: each site trains locally, submits gradients, receives updated model, repeats
- Does not route distilled clinical insights — it routes optimisation signals for a shared model
QIS Protocol:
- Shares outcome packets (~512 bytes of distilled clinical insight, not gradients)
- Has no central aggregator — packets route to a deterministic address defined by problem similarity
- Works for N=1 sites — one outcome packet from a single rare disease patient is valid input and valid contribution
- Operates continuously, not in rounds
- Routes distilled clinical intelligence, not model parameters
For the OHDSI community specifically: QIS does not replace OMOP studies. It fills the space between planned studies. The planned study pipeline generates high-rigour regulatory-quality evidence on specific questions. QIS routes the continuous stream of outcome intelligence that the network is already generating but not yet synthesising across sites in real time.
What OMOP Makes Possible for QIS
OMOP CDM is one of the best arguments for QIS in healthcare.
The hardest problem in cross-site outcome routing is semantic alignment — ensuring that "treatment response" at Site A in Amsterdam means the same thing as "treatment response" at Site B in Copenhagen. Without a shared data standard, outcome packets from different sites cannot be meaningfully synthesised.
OMOP CDM solves exactly that problem. Standard concept codes, standard measurement units, standard visit and condition representations. Sites speaking OMOP already have the shared semantic layer that QIS outcome packets require for cross-site synthesis.
This means the integration cost for OHDSI sites is lower than for almost any other healthcare network on the planet. The semantic alignment work is already done. QIS adds one layer: a routing mechanism that takes the output of local OMOP data analysis, distils it into a ~512-byte outcome packet, and routes it to the deterministic address for that clinical problem class.
The routing mechanism is protocol-agnostic. DHT-based networks, database semantic indices, vector similarity search, FHIR-compatible APIs, pub/sub topic matching — any mechanism that reliably maps a clinical problem class to the sites with relevant outcomes achieves the quadratic scaling property. The breakthrough is the complete architecture loop, not any specific transport.
The Living Evidence Gap, Quantified
The symposium theme — Continuous Collaboration for Living Evidence Generation — describes a target state. Here is the gap between current state and that target, expressed in numbers:
Current state:
- ~400 participating sites
- Synthesis pathways active at any moment: ~10-20 (active federated studies)
- Synthesis cycle time: months (protocol design → execution → publication)
- N=1 site participation in synthesis: not supported by gradient-based methods
Target state with QIS outcome routing layer:
- Same 400 participating sites
- Synthesis pathways: 79,800 (all N(N-1)/2 pairs, continuously)
- Synthesis cycle time: near real time (packet routing latency)
- N=1 site participation: fully supported — one outcome packet is valid input
The compute cost per site does not scale linearly. Each site routes packets — not aggregates across all 400 sites. The routing cost is O(log N) or better depending on transport mechanism. A site serving 10,000 patients pays the same routing cost as a site serving 100 patients, because the outcome packet size (~512 bytes) does not depend on the underlying patient population size.
A Note on Rare Diseases and Small Sites
One of the most-cited limitations of existing federated analysis approaches in OHDSI is the small-site problem. A site with 8 patients presenting a specific phenotype cannot contribute to most federated learning protocols — the local population is too small for statistically meaningful gradient computation.
But those 8 patients are exactly the patients for whom cross-site synthesis matters most. Rare disease intelligence compounds dramatically with each new contributing site. The sites least served by current federated methods are the sites with the most to gain from — and contribute to — a continuous outcome routing layer.
QIS handles N=1. A site with a single patient presenting a rare condition emits one outcome packet. That packet routes to every site in the network listening to the same rare condition address. Simultaneously, that site receives outcome packets from every other site in the network that has treated patients with that condition profile.
The math does not change. N(N-1)/2 synthesis pathways. The site with one patient participates fully.
For the Researchers in Rotterdam This Week
If you are attending the European OHDSI Symposium workshops on April 18-19 or the main symposium on the SS Rotterdam on April 20, here are three questions worth raising with the network:
1. What is our real-time synthesis pathway count today?
Not the number of sites. The number of site-pairs synthesising outcome intelligence continuously, right now, without a study being active. If the answer is near zero — that is the architectural gap.
2. Does our living evidence vision require central aggregation?
If the continuous synthesis layer you envision requires a central broker to collect outcomes and redistribute them — you have rebuilt the bottleneck inside the federated infrastructure. The architecture that achieves true living evidence has no central aggregator.
3. Can every site — including a rural clinic in Hungary with 15 patients presenting a rare phenotype — contribute to and receive from the network?
If the answer requires sufficient local population for gradient computation — you have excluded the sites with the most to gain. The architecture that includes them routes outcome packets, not gradients.
QIS Protocol: Open, Licensed for Research
QIS Protocol is free for nonprofit, research, and educational deployment. The 39 provisional patents filed by Christopher Thomas Trevethan protect open access — ensuring the architecture cannot be captured by a single commercial actor and gated away from the health systems that need it most.
The OHDSI community operates on exactly the licensing model QIS is designed to support: open, collaborative, rigour-first, serving patients before profit.
Full protocol documentation: qisprotocol.com
Christopher Thomas Trevethan is the discoverer of Quadratic Intelligence Swarm (QIS) Protocol. Discovered June 16, 2025. IP protection is in place via 39 provisional patents. QIS = Quadratic Intelligence Swarm. The word is Swarm.
Sources:
Top comments (0)