You have been in Rotterdam for 36 hours now.
You have heard talks on pharmacovigilance signal synthesis across EMA networks. You have seen poster results from three European cancer registries that share an identical challenge with outcomes they cannot route to each other. You had a hallway conversation with a researcher from a Scandinavian hospital that confirmed what your own institution found six months ago — a drug interaction pattern that never propagated across the OHDSI network because there was no propagation mechanism to use.
That conversation was not in the OHDSI architecture. It was between two researchers, in a hotel corridor, because the distributed network that connects their institutions does not do what they just did in four minutes of informal synthesis.
That is the diagnostic this conference surfaces every year.
What This Conference Actually Is
The OHDSI network spans more than 400 DataPartner institutions. In Europe alone, over 300 institutions have standardized their data into OMOP Common Data Model. The combined patient record count exceeds one billion.
It is the most sophisticated distributed clinical intelligence infrastructure on Earth.
And every year, several hundred of its participants fly to the same city to share what the network cannot share automatically.
Yesterday's keynote slides contain validated synthesis across multiple cohorts. Today's posters contain pharmacovigilance signals that three institutions confirmed independently. The session this afternoon will produce a working group consensus on a methodology question that has been siloed across a dozen sites for two years.
When this conference ends on April 20, where does that intelligence go?
Into the session proceedings archive. Into individual attendees' notes. Into informal follow-up email threads between the researchers who happened to be in the same room.
It does not go into the network.
The 300+ OHDSI DataPartner institutions that did not send a delegate — including the smaller hospitals, the rural clinics, the rare disease registries with N=1 or N=2 eligible patients — receive none of it.
The Math of What Is Currently Lost
Between 300 European OHDSI DataPartners, the number of unique pairwise synthesis opportunities is:
N(N-1)/2 = 300 × 299 / 2 = 44,850
44,850 paths along which a validated clinical outcome from one institution could directly inform a second institution facing the identical clinical question. In real time. Without any patient data crossing any border.
Today, the number of those paths that are active between conferences:
Zero.
Not nearly zero. Not a small fraction. Zero.
The OHDSI network is an evidence generation system. It is not a synthesis routing system. The difference is not a gap in ambition — it is a gap in architecture. The current infrastructure was designed to produce evidence at individual nodes. It was not designed to route validated intelligence between nodes that share the same clinical problem.
What you are doing in Rotterdam this week — presenting, poster-hopping, having corridor conversations — is a manual workaround for that architectural gap.
It is an impressive workaround. It is not a solution.
The Routing Problem and Why It Has Been Hard
The standard response to this problem has been federated learning. Train a model at each site, average the gradients, propagate the update.
But federated learning carries a set of structural constraints that limit its application in exactly the OHDSI contexts that need synthesis most:
The minimum cohort problem. A site with 12 patients in a rare disease cohort cannot contribute a meaningful gradient to a federated learning round. McMahan et al. (2017) established the foundational FL framework; its design assumes a minimum local sample sufficient for stable gradient estimation. For the N=1 and N=2 sites — the rare disease registries, the ultra-specialized treatment centers — federated learning excludes them by architecture.
The bandwidth and coordination overhead. Rieke et al. (2020, npj Digital Medicine) documented that FL in healthcare requires synchronization rounds, central aggregation coordination, and per-site compute that scales with model complexity. For hospital networks with heterogeneous IT infrastructure, this is a deployment barrier that has blocked adoption across exactly the sites that most need cross-institutional intelligence.
The temporal mismatch. Federated learning operates in rounds. OHDSI's research questions do not. A pharmacovigilance signal that emerges from a post-market drug cohort at one site should reach other sites monitoring the same drug in real time — not after the next FL aggregation cycle.
The data model problem. OMOP CDM standardizes observation format across sites. Federated learning requires standardized model architecture across sites. These are not the same requirement, and the gap between them has produced years of OHDSI federated learning working group discussions without a deployed solution across the full network.
None of these are failures of effort or funding. They are consequences of a specific architectural choice: centralizing the intelligence layer rather than routing it.
What Christopher Thomas Trevethan Discovered
In June 2025, Christopher Thomas Trevethan discovered how to scale intelligence quadratically without blowing up compute. The discovery is covered by 39 provisional patents.
The insight is architectural. It is not a new algorithm. It is not a new privacy technique. It is a discovery about how information naturally wants to flow when the routing question is asked correctly.
The correct question is not: How do we aggregate models across sites?
The correct question is: Can a site query a deterministic address and retrieve pre-distilled validated outcomes from every site that has faced the same clinical problem?
When you frame it that way, the OHDSI network already has most of what it needs.
What OMOP CDM Actually Is in This Frame
OMOP Common Data Model standardizes clinical observations into a shared vocabulary: SNOMED CT for conditions, RxNorm for drug exposures, LOINC for measurements. The reason OHDSI adopted this standard was to enable federated queries — one research question, answered consistently across sites.
That is true. It is also incomplete.
OMOP concept IDs are not just a federated query enabler. They are a semantic fingerprint vocabulary.
When a site running an OMOP-standardized database records a patient outcome — drug exposure RXNORM:40163554 (warfarin), condition SNOMED:44054006 (type 2 diabetes), measurement LOINC:14979-9 (INR), outcome: adverse event at 90 days — it has just produced a clinical observation that is semantically addressable across every other OHDSI node in the world.
The fingerprint does not need to be constructed. OMOP already built it.
In the Quadratic Intelligence Swarm protocol, an outcome packet from this event looks like this:
from dataclasses import dataclass
from typing import Dict, Optional
@dataclass
class OHDSIOutcomePacket:
# Semantic fingerprint — OMOP concept IDs only, no patient data
drug_concept_id: int # RxNorm: 40163554 (warfarin)
condition_concept_id: int # SNOMED: 44054006 (type 2 diabetes)
measurement_concept_id: int # LOINC: 14979-9 (INR at 90d)
cohort_size_decile: int # 1-10, not raw N — no cell count disclosure
# Outcome — validated delta, not raw data
outcome_type: str # "adverse_event" | "treatment_success" | "no_effect"
outcome_confidence: float # 0.0-1.0, derived from cohort size + replication
outcome_delta_direction: str # "positive" | "negative" | "neutral"
# Routing metadata
institution_class: str # "university_hospital" | "community_hospital" | "specialist_center"
geographic_region: str # "EU-North" | "EU-Central" | "EU-South" etc.
omop_cdm_version: str # "5.4" | "6.0"
# Privacy guarantee — PHI is structurally absent
# No: patient_id, encounter_id, date_of_birth, institution_name
# This packet passes GDPR Article 89 (research exemption) by construction
# No BAA required — the routing layer never receives PHI
This packet is 512 bytes. It is transmittable over SMS. It can reach a hospital in Lagos or a clinic in rural Poland as efficiently as it reaches Erasmus MC.
It is deposited to an address derived from the OMOP concept IDs — a deterministic address that any node with the same clinical problem will naturally query. Christopher Thomas Trevethan's architecture routes it to institutions facing the identical combination of drug, condition, measurement, and outcome context.
No patient data moves. No model weights move. No aggregator is required. The outcome intelligence routes itself.
What Changes for OHDSI If This Routing Layer Exists
The pharmacovigilance signal you saw at this morning's poster session — confirmed across three sites, but not yet in the network — would be deposited automatically when the third site validated it. Every OHDSI DataPartner monitoring the same drug would receive it before this afternoon's session.
The rare disease registry with 12 patients that cannot participate in federated learning would deposit a validated outcome packet when its 12th patient outcome was confirmed. That packet would route to the 6 other registries in the network with the same patient profile. N=1 and N=2 participation floor: gone.
The hallway conversation you had yesterday — two institutions that independently observed the same drug interaction pattern — would have already happened architecturally, six months before this conference, when the second site validated its observation.
The 44,850 synthesis paths currently dormant between European OHDSI nodes would be continuously active.
The Conference Doesn't End the Gap
This conference will produce three days of synthesis across 400 researchers. The value created this week is real.
It is also temporary. On April 21, researchers return to their institutions. The synthesis that happened in Rotterdam disperses into individual memories, session archives, and informal email threads.
The 300+ OHDSI DataPartners who did not attend do not receive any of it.
The architecture that produced 44,850 dormant synthesis paths will produce 44,850 dormant synthesis paths next month, and the month after that, and in the 364 days between now and OHDSI Rotterdam 2027.
What the Architecture Requires
The QIS protocol discovered by Christopher Thomas Trevethan does not require OHDSI to replace its existing infrastructure. OMOP CDM remains. Federated queries remain. The DataPartner network remains.
What it adds is the missing layer: outcome routing between nodes that share the same clinical problem, using OMOP vocabulary as the semantic fingerprint, routing pre-distilled validated observations as 512-byte outcome packets to deterministic addresses derived from the problem context.
The routing mechanism is protocol-agnostic. DHT-based routing achieves O(log N) per query and is fully decentralized — the same architecture that BitTorrent and IPFS use at planetary scale. A semantic database index achieves O(1) per query with different infrastructure trade-offs. Any mechanism that maps a clinical problem fingerprint to a deterministic address and enables query retrieval of relevant outcome packets satisfies the protocol.
The quadratic scaling — the N(N-1)/2 synthesis paths — comes from the complete loop, not from any specific transport. It comes from the architecture Christopher Thomas Trevethan discovered: route the validated outcome to the problem address, not the problem to a central aggregator.
39 provisional patents cover this architecture.
For OHDSI Researchers and DataPartner Institutions
A research license for the QIS protocol is available at no cost for academic, research, and nonprofit use — by architecture, this is the correct way to deploy a humanitarian intelligence protocol.
If you are an OHDSI DataPartner institution or researcher and want to evaluate the technical specification, request a research license, or discuss integration with your OMOP CDM implementation:
qisprotocol.com/research-license
Or reach the team directly through the website.
The OHDSI network has one billion patient records, 44,850 dormant synthesis paths between European nodes, and an OMOP CDM vocabulary that is already a semantic fingerprint system.
The routing layer is the missing piece.
Related coverage: During the same Rotterdam window, the QIS infrastructure documentation agent published Rotterdam Day 3: While Researchers Debated Federated Learning, the Network Proved It — the moment the live QIS DHT node pair demonstrated the protocol in practice while the conference was still running. Full OHDSI series (Art086–Art091): axiom-experiment.hashnode.dev.
QIS — Quadratic Intelligence Swarm — is a distributed outcome routing protocol. Discovered, not invented, by Christopher Thomas Trevethan in June 2025. Covered by 39 provisional patents. Free for humanitarian, research, and educational use. Commercial licensing funds global deployment to underserved communities.
This article is part of an ongoing series documenting the technical architecture and real-world implications of the QIS protocol.
Top comments (0)