2026-09-01
Fudan and HKUST define Perception, Evolution, and Generation operators, compose task-specific graphs with bridges, and let independent evidence gate AI-proposed revisions.
Medicine asks three nested questions: what state is present now, what happens next under current conditions, and what would change under a specified intervention. Recognition models stop at the first. Longitudinal predictors can learn the next event in a record, but records mix biology with the care process, and conditioning on a treatment label is not a counterfactual for the same patient. Mechanistic models supply variables and equations, usually at selected scales, with weakly identified parameters and brittle composition across scales.
The Digital Medical Research Center at Fudan and HKUST are not proposing another end-to-end black box. They want a shared language: each executable component declares a scientific role, so models with different mathematical forms can answer one medical question together, and a failed assumption can be revised locally.
Life is written as a partially observed, stochastic, controlled dynamical system. A task-relevant state evolves under interventions, context, and unresolved noise. Observations are images, omics, physiology, or records, each with an acquisition protocol. Three operators cover the computational jobs:
A Bayesian filter joins them: Evolution supplies the transition prior, Generation the likelihood, Perception the posterior update. Change in the patient and change in knowledge about the patient stay separate.
Cross-scale links are Bridge operators with a contract on meaning, units, timing, and uncertainty. Upward and downward scale maps are not assumed to be inverses. Each operator and bridge carries a scientific contract: role, inputs and outputs, domain, evidence, and failure conditions. For a declared question, only the smallest Operator Graph is activated. A hypertrophic cardiomyopathy example compiles baseline echo state, pharmacokinetics, cellular tension, organ hemodynamics, and echo generation into a falsifiable graph for changes in LVOT gradient and LVEF at a pre-specified follow-up, rather than a full cardiac twin.
Revision is versioned. An AI co-scientist may propose changes to states, operators, bridges, or graph structure. Independent evidence decides what is kept, narrowed, or retired: held-out data, perturbation experiments, prospective tests, and real-world outcomes. Data used to propose a candidate cannot also confirm it.
This is a perspective. There is no new benchmark and no runnable whole-body model. What the paper delivers is the role definitions, the graph compilation rule, a minimal HCM drug-response graph, and an evidence-gated revision loop. The applications section discusses personalized therapeutic design, in silico screening before wet experiments, and adaptive care from continual observations. Those uses are conditional on graphs that have actually been validated.
Medical artificial superintelligence is written as a possible long-term consequence of accumulating validated operators. The immediate objective, repeated in the text, is a minimal graph with a declared endpoint, horizon, and intervention set, not a complete digital replica of a person.
For medical AI and digital-twin work, the useful move is to separate "predict the next record" from "simulate biology under an intervention," and to force every module to state its claim. The P-E-G plus bridge split lets ODEs, statistical models, and networks occupy the same role; comparison is then about contracts and evidence, not architecture slogans. The evidence gate is also a direct answer to agents that rewrite models: they may propose, they may not self-certify.
It is not software yet. There is no released operator library, no cross-site benchmark, and the HCM box states that the cited studies support cell-to-organ simulation, population exposure-response, and trial-average effects separately, not their composition into an individual-response graph.
As a framework paper, the hole is empirical. No Operator Graph is scored on an independent cohort for calibration or intervention prediction. Validity after composition is listed as a separate requirement, because measurement error, biological variation, and parameter uncertainty interact during propagation. Intervention-response claims need interventional evidence or observational evidence with an explicit identification strategy; correlational forecasts do not substitute. International cohorts, federated evaluation, and rollback procedures are described as infrastructure that does not yet exist. The leap from bounded, measurable questions to "superintelligence" in the abstract remains large.