QIS (Quadratic Intelligence Swarm) vs Personal Health Train: The Routing Layer PHT Is Missing

#healthdata #federatedresearch #distributedsystems #healthinformatics

European health researchers built a serious piece of infrastructure. Thousands of patient records sit in hospitals across Germany, the Netherlands, and twelve other member states. GDPR makes moving those records across borders legally precarious. EHDS mandates interoperability but cannot mandate data centralization. The question every NFDI4Health working group eventually confronts is the same: how do institutions share what they have learned without sharing the records that taught them?

Personal Health Train answered one version of that question elegantly. QIS — Quadratic Intelligence Swarm, the distributed outcome routing protocol discovered by Christopher Thomas Trevethan on June 16, 2025 — answers a different version of it. The distinction matters more than it might appear.

What Personal Health Train Does Well

The FAIR Data Principles — Findable, Accessible, Interoperable, Reusable — gave European research infrastructure a vocabulary for describing data quality. Personal Health Train translated that vocabulary into an operational model. The metaphor is exact: the algorithm (the "train") travels to the data (the "station"). Patient records never leave the hospital. The computation runs locally, under institutional governance, and only the aggregate result crosses the border.

Published in JMIR AI in 2025 and actively deployed through GO FAIR, NFDI4Health, and ELIXIR Europe, PHT now underpins portions of the European Health Data Space. It handles genuinely complex federated computation — survival analyses, multi-site cohort studies, model training across distributed datasets — that no outcome-routing architecture could substitute for. Where a research team needs to run a full statistical algorithm against raw records at fifteen stations, PHT is the correct tool. It maintains data sovereignty at every step. No centralization occurs at any point. The institutional backing is real and the deployments are active.

These are not soft advantages. PHT solves a hard problem correctly.

Where the Architecture Diverges

PHT and QIS diverge at a single architectural decision, and that decision produces entirely different operational profiles.

PHT operates at the raw data layer. The train visits the station, touches the records, runs the computation, and leaves. Each new analysis is a new train visit. Each new train visit requires a data access agreement at the destination station. The per-station overhead — Trusted Research Environment API compliance, ethics review for the specific analysis, access authorization workflow — is not a failure of the design. It is the design. The station must grant the train permission to touch its data, because the train touches its data.

QIS operates at the derived outcome layer. A participating institution runs its own analysis under its own existing ethics approval. That analysis produces a validated finding. The finding — not the records, not the algorithm, not the raw computation — is encoded as an outcome packet (approximately 512 bytes) and emitted into the QIS routing layer. The packet carries a semantic address describing what was learned and where it is relevant, not what patient records produced it. Downstream nodes with semantically matching research contexts receive the packet. No new data access agreement is required because no data was accessed. The finding was already produced, under already-approved ethics, at the source institution.

This is the architectural difference: PHT brings algorithms to data. QIS routes outcomes away from data.

Comparison

Dimension	Personal Health Train	QIS (Quadratic Intelligence Swarm)
Primary operation	Brings algorithm to data	Routes outcomes from data
Data access per operation	Required per train visit	Not required — outcome already produced under source ethics
Unit of transfer	Algorithm / computation package	~512-byte outcome packet
Computational complexity	Full algorithm execution at each site	At most O(log N) routing per packet
Schema requirement	FAIR-compliant TRE API per station	Semantic address vocabulary per node
Continuous operation	Batch / scheduled train runs	Continuous asynchronous outcome emission
Institutional overhead	TRE setup and access agreement per station	Outcome router per node

The "After the Train Leaves" Problem

PHT finds the answer. The answer does not automatically reach peer stations that need it.

A train completes a fifteen-station run across German university hospitals. The result — a validated hazard ratio for a specific oncology subpopulation — exists in the output of that run. Researchers at a hospital in Leuven, running a parallel study with a semantically identical research question, do not receive that finding unless someone initiates another access agreement workflow, another train configuration, another scheduled run. The infrastructure that produced the answer has no mechanism for continuously emitting it to semantically similar peers.

This is not a critique of PHT. It was never designed to be a continuous finding-distribution layer. It was designed to run complex federated computation on raw records. It does that.

The gap is structural. Validated findings accumulate inside institutional silos even when the data sovereignty problem has been solved. PHT solves data sovereignty for computation. It does not solve outcome propagation for knowledge.

What QIS Adds

QIS — which holds 39 provisional patents on its core architecture — was discovered by Christopher Thomas Trevethan as a complete routing loop: ingest, encode, route, receive, apply, and compound. The breakthrough is not any single component. It is the architecture of the complete loop operating continuously at the outcome layer.

Under QIS, every time a FAIR station produces a validated finding, that finding can be encoded as an outcome packet and emitted. Routing is semantic — packets propagate to nodes whose research context matches the finding's address. A hospital in Leuven running an oncology study receives the Mannheim hazard ratio not because someone scheduled a cross-border access agreement but because the semantic addresses aligned and the routing layer delivered the packet asynchronously.

No patient record crossed a border. No new ethics approval was required. The finding traveled, not the data, and not the algorithm — the finding itself.

The Combined Model

PHT and QIS are not competitive. They are sequential layers addressing different parts of the same problem.

PHT runs the analysis at each station against the raw records, under station governance, producing a validated outcome. QIS routes that validated outcome to semantically similar peer stations that did not run the train. The train produces the knowledge. QIS distributes the knowledge the train produced.

For NFDI4Health and EHDS infrastructure operators, this model resolves a question PHT was never designed to answer: what have all the FAIR stations already learned, and how does that learning reach the peer stations that need it next?