2026-09-28
HealthFlux learns a 384-d health state from 502k UK Biobank people; five-year AUROC on 195 diseases plus death is 0.816 vs 0.715 for Delphi-2M.
Clinical events, labs, proteins, metabolites, genotypes, and MRI all describe one person, at mismatched times and in mismatched units. Most disease models still answer a single endpoint from a single modality. They do not maintain a health state that can be rolled forward and revised. A world model, in this paper, means exactly that: a latent that evolves between visits and jumps when new evidence arrives, from which risk for any disease can be read at any age.
UK Biobank supplies the scale: 502,166 participants, 5,647 features, eleven domains. Delphi-2M already learned disease natural history with a generative Transformer, but with thinner modality coverage and no continuous-time revision. HealthFlux treats health itself as the object to simulate. The corresponding authors are at Stanford (Wyss-Coray and Ying), with collaborators at the University of Washington, Harvard Medical School, and MIT CSAIL.
The backbone is a hybrid continuous-time recurrent state-space model. Between observations, an age- and context-conditioned neural ODE advances a 256-d deterministic state with fixed-step RK4, time normalized by 3652.5 days, at most four substeps per interval. A new observation defines a 128-d diagonal-Gaussian prior and posterior; a GRU then performs a gated jump. The readout state is a layer-normalized 384-d concatenation. With no observation, the ODE value is kept, so querying an arbitrary age does not pretend an unmeasured assay was taken. Explicit age encodings let the same calendar gap induce different slopes at different life stages.
The world model is not trained on incident-disease labels. It predicts the next EHR token and a transformed time gap, plus a KL term weighted 0.01. Numeric modalities update the state but are not next-event targets. Disease heads are fit later on frozen states. Training uses 351,166 people, at most 96 observation nodes per window, global batch 3,072 on six GPUs, for 2,000 steps. Only the EHR embedding table is warm-started; dynamics and multimodal encoders start from scratch.
Inputs span longitudinal records, demographics, genetics, biochemistry, haematology, physical function, lifestyle, proteomics, metabolomics, abdominal MRI, and brain MRI: 5,604 time-varying features plus 43 permanent context variables. Numeric modalities present at the same age are pooled with softmax attention; EHR tokens stay on their own nodes and precede same-age numeric rows.
On held-out UKB, 195 diseases plus death, five-year window: mean AUROC 0.816 and AUPRC 0.079. The strongest comparator is TabPFN-v3 at 0.791 / 0.063. Delphi-2M is 0.715 / 0.037; XGBoost 0.790 / 0.064; RisQ 0.790 / 0.064. Clinical scores lose on their own endpoints: composite CVD 0.776 vs QRISK3 0.743; all-cause death 0.846 vs Charlson 0.754.
When target diagnoses are removed from both world-model and readout training, strictly unseen diseases still reach mean AUROC 0.769. Eight held-out demographic groups land at 0.755, above XGBoost at 0.734. Simulated forward with no further observations, the AUROC edge over Delphi-2M grows from 0.098 at a one-year gap to 0.128 at five years, so the last window is years 5–10 after the last measurement. Top-10% five-year incidence enrichment across ICD chapters is about 4.0–9.7×; top-1% is about 6.6–62.3×. People with six or more new diagnoses in 180 days change state 4.4× faster than people with none.
External cohorts use label-free test-time training that only updates normalization: median disease-level AUROC is 0.852 in CARDIA, 0.670 in ARIC, and 0.604 in WHI, each above Delphi-2M and RisQ on the shared evaluable set. For hypertension, the pan-modal model recovers 27.9–55.5% of cases missed by single-modality models at a 90% specificity threshold.
This is a serious attempt to put a world-model loop (latent, dynamics, observation update) onto pan-modal health data. Because the disease head is fitted later, rare conditions, new cohorts, and new assays have a transfer path. Continuous time also matches the data better than treating every visit as a discrete token: measurements are sparse, disease is not.
It is not a clinical tool. The preprint says so. UK Biobank healthy-volunteer bias and recorded incidence rather than true onset both push numbers up. AUPRC 0.079 means absolute precision on rare events is still low; a strong AUROC does not make a screening test.
External prediction depends on unlabeled adaptation. Several transfer and biology figures use an earlier continuous-time checkpoint and task-specific heads, so the main model and the analysis model are not the same weights. WHI’s median AUROC is only 0.604; cross-cohort behavior is uneven. The ODE step cap and the 96-node window drop long histories. Unseen-disease 0.769 is a mean; whole-chapter holdouts drop, more so as the transferred head sits farther away (Pearson r=−0.278). There is no prospective intervention study, so counterfactual “what if we ordered this test” is untested. Two thousand training steps is short; convergence of the world model is thinly documented.