Multimodal health world model that evolves a latent state from records, blood tests, proteomics, metabolomics and MRI to forecast 195 diseases.
No providers recorded yet. Browse all providers
A patient's health is never observed directly. A blood panel captures one facet in March, a brain MRI another three years later, and a diagnosis enters the record only once a disease is clinically recognized. Most disease-risk models take a snapshot of one visit or read a sequence of clinical codes, so they cannot say what happened between measurements or fold a new proteomic profile into what they already know. HealthFlux treats health as a single hidden state that evolves continuously in time, is simulated forward between observations, and is revised whenever any new measurement arrives, which is the defining structure of a world model.
HealthFlux was developed by Stanford University in the Wyss-Coray lab, with collaborators at the University of Washington, Brigham and Women's Hospital and Harvard Medical School, MIT CSAIL, the University of Pennsylvania, Karolinska Institutet and the University of Oxford; Kejun Ying and Tony Wyss-Coray are co-corresponding authors. It was posted to medRxiv on 21 September 2026 and has not yet been peer reviewed.
Where EHR foundation models such as CEHR-GPT and the Verily EHR and genomics model learn from sequences of clinical events, HealthFlux also ingests plasma proteins, metabolites and MRI-derived anatomy.
HealthFlux is a hybrid continuous-time recurrent state-space model. Permanent context (40 genetic principal components, birth year, sex and ethnicity) is encoded into a 128-dimensional vector, and attained age into a 64-dimensional one. A 256-dimensional deterministic state is integrated with fourth-order Runge–Kutta between observations; at each observation a 128-dimensional diagonal-Gaussian stochastic state is inferred, and the two are concatenated into a 384-dimensional latent state. The dynamic record holds 1,257 clinical-event identities and 4,347 numeric features. The world model is trained self-supervised to predict the identity and timing of the next clinical event, with a KL term between prior and posterior; disease labels are not used to fit it. Training used 351,166 of 502,166 UK Biobank participants, on six GPUs for 2,000 steps at a global batch size of 3,072.
On 75,318 held-out UK Biobank participants, five-year prediction of 195 diseases and death reached a mean AUROC of 0.816, against 0.715 for Delphi-2M and 0.791 for TabPFN-v3, the strongest comparator; XGBoost, LightGBM and RisQ were also evaluated. With all subsequent observations withheld, its margin over Delphi-2M grew from 0.098 at a one-year gap to 0.128 at five years. In three US cohorts (CARDIA, ARIC and WHI), a label-free test-time adaptation of layer-normalization parameters gave median disease-level AUROCs of 0.852, 0.670 and 0.604. When target diagnoses were excluded from both readout fitting and world-model training, disease transfer reached a mean AUROC of 0.769; the transfer and modality analyses used an earlier continuous-time implementation of the model.
HealthFlux is aimed at population-health and ageing researchers who want to estimate risk across many diseases at once from whatever data a person has, from a hospital record alone to one with proteomics and imaging. Its forward simulation supports long-horizon risk estimation when follow-up data are sparse, and its modality attributions show which measurements drive each prediction: blood biochemistry and metabolomics for metabolic disease, proteomics for skin disease, and abdominal MRI for musculoskeletal disease. The authors also use it to identify future hypertension cases that single-modality models miss.
HealthFlux moves disease-trajectory modelling beyond sequences of clinical codes to a continuous-time, multimodal world model, and its strict-exclusion experiments test whether a learned health state carries information about diseases it never saw. All results come from UK Biobank and three US cohorts and are retrospective; the authors note that external prediction relied on test-time adaptation and that several interpretation analyses used earlier model versions. The preprint includes no code or model-weight release, and the underlying participant data are available only through each cohort's access process, so the model cannot currently be reproduced or applied by outside groups.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.