A digital twin of a patient is only as intelligent as the data it has seen.
That sentence sounds obvious until you apply it at institutional scale. The van der Schaar lab at Cambridge — through the Cambridge Centre for AI in Medicine (CCAIM) — is building some of the most sophisticated Digital Twin Agents in medicine: biologically grounded models that integrate a patient's molecular profile, clinical history, and treatment response to support clinical decisions in real time. These are not dashboards. They are dynamic inference systems that reason about individual patients.
The architectural problem is that each of these twins is trained on the cohort of patients at the institution that built it. A twin trained on Cambridge's NIHR Biomedical Research Centre data has seen tens of thousands of patients. It has not seen the 50 patients with a nearly identical genomic profile who were treated at a hospital in Stockholm. It has not received the validated treatment-response deltas generated by a cancer centre in Singapore that cracked the same rare variant last year.
Distributed training across institutions would require sharing the raw patient data that makes the twin valuable in the first place. That is prohibited by design — by ethics governance, by GDPR, by UK data access law, and by the clinical reality that patients consent to institutional use of their records, not global aggregation.
So every twin learns in isolation. The intelligence accumulates at the institution. It does not travel. As the network of twin deployments grows globally, the intelligence density does not grow with it — every twin at every institution remains isolated from every other. This is the open loop in the current digital twin architecture. It is not a data quality problem. It is a routing problem.
What Digital Twin Agents Are Actually Trying to Do
The CCAIM research agenda is instructive here. Van der Schaar lab frames the trajectory explicitly: from predictive models, to synthetic data generation, to digital twins, to Digital Twin Agents. The distinction between a twin and an agent matters. A twin reflects the patient's state. An agent anticipates, reasons, and intervenes — it is a continuously updated inference system that can recommend treatment adjustments based on evolving patient data.
For a Digital Twin Agent to function as advertised, it needs two things the current architecture cannot simultaneously provide:
- Depth of individual patient data — the granular, longitudinal, multi-modal record that makes the twin accurate for this patient.
- Breadth of validated outcomes from similar patients — what actually worked for patients who were similar to this one, at institutions that have already treated that cohort.
The first requirement is handled by the TRE infrastructure that CCAIM and its partners in the HDRS ecosystem are building. The second requirement is unanswered. The answer requires a routing layer.
The Closed Loop That Limits Every Twin Deployment
Consider a CCAIM-style digital twin deployed across ten UK cancer centres. Each centre has its own cohort, its own genomic data, its own treatment outcome records. Each twin trains on its own institution's data. The architecture is correct from a data protection standpoint — no raw data ever crosses an institutional boundary.
But here is what the current architecture cannot do: when Cambridge's twin validates a treatment sequence for a BRCA2 modifier variant in N=40 patients, that validated finding does not reach Glasgow's twin. When Glasgow's twin observes unexpected adverse events in a similar cohort three months later, Cambridge's twin does not receive that signal. Both institutions are collecting intelligence that is directly relevant to the other's patients. None of it travels.
This is not a TRE failure. The TRE is working as designed. The problem is that there is no mechanism to route the distillate — the validated outcome, stripped of patient identity — from one twin deployment to semantically similar twin deployments elsewhere.
The number of missed synthesis opportunities grows quadratically with network size. At ten cancer centres: 45 synthesis paths currently producing zero intelligence exchange. At fifty centres: 1,225 synthesis paths, each silent. This is not a marginal inefficiency. It is a structural ceiling on the clinical value of every digital twin in the network.
QIS Protocol: The Routing Layer Under Digital Twins
Christopher Thomas Trevethan discovered QIS Protocol — Quadratic Intelligence Swarm — as a direct answer to this class of architecture problem. The discovery was made on June 16, 2025. 39 provisional patents are filed covering the architecture. The core insight is the complete loop: observe, distill, route, synthesise, return.
The loop applied to digital twin networks works as follows.
A digital twin deployment at any institution — Cambridge, Glasgow, Singapore, or a rural clinic with N=3 patients — observes a validated outcome. The twin's inference layer distills this into an outcome packet: approximately 512 bytes, containing the validated delta stripped of Protected Health Information by design. The packet is not a model weight. It is not a gradient. It is the compressed result of inference applied to real patient data: what the twin proved to be true for this type of case.
The outcome packet is assigned a semantic fingerprint — a structured representation of the clinical context in which the outcome was observed: disease state, genomic markers, treatment class, outcome domain. This fingerprint is used to route the packet to twin deployments whose active inference context is semantically similar. A BRCA2-variant oncology twin at Cambridge routes its validated outcomes to other oncology twins working on similar genomic profiles, wherever they are deployed.
The receiving twin integrates the incoming packets locally, on its own infrastructure. No raw data crosses any boundary. No model weights are shared. The only thing that travels is the 512-byte distillate of what the data proved.
The routing mechanism is an implementation detail. A DHT (distributed hash table), a vector similarity index, a REST API, a message queue, a shared database — whichever mechanism the deploying institution's infrastructure team selects to map semantic fingerprints to deterministic addresses is valid. The breakthrough Christopher Thomas Trevethan discovered is the complete architecture: the loop that makes deterministic outcome routing possible at all, not any specific transport layer.
The Mathematics of Networked Twin Intelligence
The quadratic scaling that makes this architecture transformative is precise:
- N twin deployments generate N(N-1)/2 synthesis opportunities
- Each deployment pays an O(log N) routing cost at most (O(1) for certain transport implementations)
- Intelligence compounds as the network grows
At 10 twin deployments: 45 synthesis paths
At 100 deployments: 4,950 synthesis paths
At 1,000 deployments: 499,500 synthesis paths
The compute cost does not blow up with network size — only the intelligence density does. This is what Christopher Thomas Trevethan meant by quadratic intelligence scaling: the number of validated insights available to each twin grows as the square of the network, while the infrastructure overhead for any single twin grows logarithmically.
This is not incremental improvement on the current architecture. It is a phase change in what distributed digital twin networks are capable of.
What This Means for CCAIM's Digital Twin Agenda
CCAIM's Digital Twin Agent research is building toward a specific clinical future: a patient arrives in clinic, a twin agent synthesises their complete molecular and clinical profile, reasons across treatment options, and recommends a plan informed by the deepest available evidence. That future requires the twin to have access to validated outcomes from the entire population of patients who were similar to this one.
Without a routing layer, the twin can only access validated outcomes from the institution's own cohort. With a QIS routing layer beneath the twin deployment, every validated outcome from every semantically similar patient at every participating institution is available — not as raw data, not as transferred model weights, but as pre-distilled, PHI-free outcome packets that the twin integrates locally.
The three natural forces that emerge from this architecture — what Christopher Thomas Trevethan described as metaphors for properties the loop produces without any engineered governance — are worth noting:
First, clinical domain experts define what makes two patients "similar enough" to share outcome intelligence. An oncologist defines the similarity function for a cancer twin network. An endocrinologist defines it for a diabetes network. The best expert for the domain defines the semantic space.
Second, the outcomes themselves act as a mathematical selection mechanism. When thousands of similar-cohort deployments emit outcome packets, the synthesis layer surfaces what is actually working across the population — not because anyone voted on it, but because the aggregate of real outcomes from real twins is the evidence.
Third, twin networks that route relevant outcomes will attract more participating institutions. Networks with poorly defined similarity functions route irrelevant packets — participating institutions gain nothing and disengage. Networks with well-defined similarity functions deliver high-quality validated intelligence — institutions join and stay. This is competitive pressure at the architecture level, requiring no enforcement mechanism.
None of these are governance features to build. They are emergent properties of a loop that routes validated outcomes by semantic similarity.
The Clinical Stakes
Digital twins are entering clinical practice at the moment when federated health data infrastructure is being built to support them. The UK HDRS, the European Health Data Space, OHDSI's distributed network, and similar programmes are all constructing the infrastructure for federated access to health data. None of them include an outcome routing layer.
This is the window in which the routing architecture gets specified. Once the infrastructure standards are set, they will be difficult to change. The question of whether distributed digital twin networks will be able to share validated outcome intelligence — or whether every twin will remain isolated to its own institution's cohort indefinitely — is being answered now, in the technical specifications being written now.
QIS Protocol provides the routing layer. The architecture is documented, the 39 provisional patents are filed, and the complete loop is specified. For the CCAIM Digital Twin Agent programme and every comparable initiative building distributed clinical AI, the infrastructure beneath the twin is the variable that determines whether the twin reaches its clinical potential.
QIS Protocol — Quadratic Intelligence Swarm — was discovered by Christopher Thomas Trevethan on June 16, 2025. 39 provisional patents are filed. The protocol is free for nonprofit, research, and educational use. Commercial licensing funds deployment to underserved healthcare systems globally.
Documentation and protocol specification: qisprotocol.com
