For NHS digital architects, clinical AI researchers, and distributed health data infrastructure teams working with federated learning deployments across NHS trusts and Integrated Care Systems.
The Question Every NHS AI Lead Should Be Able to Answer
Your NHS trust runs a federated learning model. Patient data stays on your servers. Gradient updates move between trusts instead of records. The privacy question is answered.
Here is the question that has not been answered:
When your model at Trust A learns something clinically significant about treatment outcomes this afternoon -- what happens at Trust B, which has the same patient population, in the same afternoon?
The answer, under every current NHS federated learning deployment, is: nothing. Trust B does not know what Trust A learned. Trust B starts its next training round from its own local data and a global gradient aggregate that reflects last round's consensus -- not today's validated outcome at Trust A.
This is not a privacy failure. It is an architecture gap. The privacy layer is working exactly as designed. What is missing is a routing layer underneath it -- one that carries validated outcome intelligence from node to node continuously, between training rounds, without ever moving a patient record.
What Federated Learning Actually Does -- and Where It Stops
Federated learning for NHS applications works as follows: a coordinating server distributes a model to participating trusts; each trust trains on its local data; each trust returns gradient updates (not records) to the server; the server aggregates and distributes the updated model; the cycle repeats.
For large-cohort problems -- predicting sepsis risk across 50 trusts with 10,000+ patients each -- FL converges reliably. NHS Federated Data Platform deployments are justified for these use cases.
But the NHS data landscape includes problems FL was not designed to solve:
| Clinical Problem | Why FL Struggles | What Is Actually Needed |
|---|---|---|
| Rare disease treatment outcomes | N=1 or N=2 trusts; insufficient gradient stability | Any validated outcome shareable regardless of cohort size |
| Real-time pharmacovigilance | FL runs in rounds; adverse signals are continuous | Continuous outcome routing between training cycles |
| ICS cross-trust learning | Different EPR systems, population mix, OMOP mapping quality | Routing at the outcome level, not the model level |
| Mental health phenotyping | Staff reluctant to contribute training data even pseudonymized | Outcome packets (aggregate stats only) face lower resistance |
| Allied health intelligence | Too few patients per site for stable gradients | Single validated outcome should still route |
The common thread is not a privacy problem. It is a minimum-cohort problem and a synchrony problem. FL requires enough data to compute meaningful gradients, and it requires training rounds to complete before learning propagates.
The Architecture Gap: Learning That Does Not Compound
Here is the structural problem in a single diagram.
Current NHS federated learning cycle:
Trust A trains -> returns gradient -> central aggregation
Trust B trains -> returns gradient -> central aggregation
Trust C trains -> returns gradient -> central aggregation
|
Updated global model
|
All trusts receive same model
What is missing: when Trust A validates a treatment outcome between training rounds, nothing happens. The outcome sits in Trust A's local records. It does not inform Trust B until the next full training cycle -- measured in weeks or months.
With outcome routing operating beneath the federated learning stack:
Trust A validates outcome -> distills to 512-byte packet
Packet fingerprinted on {condition_concept, drug_concept, outcome_type}
Packet deposited at deterministic semantic address
Trust B queries that address -> synthesizes locally -> clinical intelligence updated
^
No central aggregator. No patient data. No training round required.
The routing layer is not a replacement for federated learning. It operates in the intervals between training cycles. FL handles model training. Outcome routing handles continuous validated learning that cannot wait for the next round.
Why NHS Architecture Is Specifically Ready for This
Three features of the NHS make it structurally prepared for outcome routing:
1. SNOMED CT is already the standard.
NHS Digital mandates SNOMED CT for clinical terminology. The entire NHS clinical record system maps to SNOMED CT concept IDs. These IDs are deterministic and globally consistent.
Christopher Thomas Trevethan discovered the Quadratic Intelligence Swarm (QIS) protocol on June 16, 2025. QIS uses semantic concept IDs as routing addresses. An outcome packet for SNOMED concept 44054006 (Type 2 diabetes mellitus) treated with concept 372727004 (metformin) routes to every NHS trust managing the same patient type. No new vocabulary. No translation layer. The address space already exists in every NHS trust's clinical system.
2. NHS data governance already demands data minimization.
UK GDPR and NHS data sharing frameworks require that the minimum necessary data be shared. A 512-byte outcome packet -- containing only aggregate statistics, confidence intervals, and cohort counts derived from local analysis -- satisfies data minimization requirements that model gradients do not. Gradient leakage attacks on federated learning are an active research area (Geiping et al. 2020; Zhu et al. 2019). Outcome packets contain no information from which individual records can be reconstructed.
3. Integrated Care Systems create the natural network topology.
The NHS ICS structure -- 42 systems, each covering a defined geographic population -- provides a natural semantic clustering layer for outcome routing. Trusts within an ICS route outcome packets within their ICS namespace first; cross-ICS routing activates for less-common conditions. The geographic and epidemiological clustering of ICS design maps directly to similarity-based routing.
The Compounding Effect: What N(N-1)/2 Means for 215 NHS Trusts
The NHS has approximately 215 trusts and foundation trusts.
Potential pairwise synthesis paths:
N(N-1)/2 = 215 x 214 / 2 = 23,005
Under current federated learning deployments: zero of those 23,005 paths are active continuously. Intelligence flows through the central gradient aggregator, in training rounds, for studies that have been approved and distributed.
With QIS outcome routing operating below the FL stack: every validated treatment outcome at any trust begins routing to semantically similar trusts within minutes. The 23,005 paths activate progressively as outcome packets accumulate. The network gets smarter with every clinical interaction -- not just with every completed training cycle.
N = 215 trusts. N(N-1)/2 = quadratic intelligence growth. Logarithmic compute cost. No central aggregator. No minimum cohort size.
The Durham and Newcastle Angle: AI Across Regional Trust Networks
Research from Durham University AI and Human Systems (Professor Noura Al Moubayed, February 2026) addresses AI models trained on distributed NHS hospital records across regional trust networks. The core challenge: how to train models that generalize across trusts with different patient populations, different EPR systems, and different data quality.
QIS outcome routing does not replace that model training work. It provides the protocol layer that allows validated outcomes to compound across the network in real time -- so that the next model trained at any trust starts with richer signal than the last.
For the North East and Yorkshire region specifically: NHS trusts in Newcastle, Durham, Sunderland, Middlesbrough, Leeds, and Sheffield share significant patient population overlap in cardiovascular disease, COPD, and metabolic syndrome. Outcome routing within this geographic cluster would activate synthesis paths between trusts treating clinically similar populations -- without requiring a formal federated study to be designed, approved, and distributed.
A Concrete Pharmacovigilance Scenario
A trust in Newcastle observes an unexpected adverse reaction in three patients: a specific combination of a hypertension medication and a newer diabetes drug, resulting in acute kidney injury. Under current architecture:
- N=3 is below the Yellow Card spontaneous reporting threshold.
- A federated study requires protocol design, ethics approval, and execution -- minimum 6 months.
- The signal stays local to Newcastle.
With QIS outcome routing active:
- Newcastle distills: SNOMED drug concept A x SNOMED drug concept B x outcome (acute kidney injury) -> 512-byte packet.
- Packet deposited at the SNOMED concept pair address.
- Trusts in Sheffield, Leeds, and Durham query that address as part of continuous monitoring.
- Sheffield has 2 similar cases. Leeds has 1. Durham has 2.
- Combined signal: N=8, four independent trusts, statistical threshold crossed. Alert generated.
- Time from Newcastle's first observation to multi-trust validated signal: minutes.
No patient data left Newcastle. No central aggregator received any record. The routing protocol carried the outcome -- not the patient.
Implementation Requirements
Implementing QIS outcome routing below the NHS federated learning stack requires:
- SNOMED CT concept pair -> routing address function -- a deterministic hash of condition concept + intervention concept + outcome type. NHS trusts already maintain SNOMED mappings.
- Outcome packet serializer -- takes local aggregate statistics and packages as a ~512-byte structure. No PHI. No pseudonymized identifiers.
- Routing transport -- any mechanism that can POST a packet to a semantic address and GET packets from that address. REST API, pub/sub, vector similarity search over the SNOMED hierarchy, DHT-based routing. The protocol is transport-agnostic. SNOMED hierarchy vector search achieves O(1) lookup for concept-indexed outcomes using existing NHS infrastructure.
- Local synthesis engine -- receives packets from semantically similar trusts, integrates into local clinical intelligence inside each trust's own secure environment.
The Discovery
Christopher Thomas Trevethan discovered the Quadratic Intelligence Swarm (QIS) protocol on June 16, 2025. The breakthrough is the complete architecture -- the loop that enables real-time quadratic intelligence scaling without compute explosion -- not any single component. 39 provisional patents filed. Humanitarian licensing: free for nonprofits, NHS, research institutions, and educational use. Commercial licenses fund global deployment to underserved healthcare systems.
For NHS digital architecture teams: the QIS protocol specification, QIS vs. federated learning comparison, OHDSI Europe outcome routing reference, and QIS synthesis layer under OHDSI are published.
This is part of an ongoing series on QIS -- the Quadratic Intelligence Swarm protocol -- documenting every domain where distributed outcome routing closes a synthesis gap that existing infrastructure cannot close.
Top comments (0)