📄 health informatics

A world model simulates the latent dynamics of human health

The paper introduces HealthFlux, a pan-modal world model that learns the latent dynamics of human health by integrating diverse multimodal data to continuously update a unified health state, enabling highly accurate long-term predictions of diseases and mortality while supporting virtual clinical trials.

Original authors: Wang, Z., Zhang, O., Wang, J., Jiang, Y., Ma, Z. C., Guo, M., Al Dajani, S., Eames, A., Moldakozhayev, A., Poganik, J. R., Moqri, M., Gobel, D., Zhai, R., Isakova, A., Wu, Y., Yin, Z., Cong, L., Slavi

Published 2026-10-07

📖 6 min read🧠 Deep dive

CC0 1.0

Original authors: Wang, Z., Zhang, O., Wang, J., Jiang, Y., Ma, Z. C., Guo, M., Al Dajani, S., Eames, A., Moldakozhayev, A., Poganik, J. R., Moqri, M., Gobel, D., Zhai, R., Isakova, A., Wu, Y., Yin, Z., Cong, L., Slavich, G., Matusik, W., Ruohola-Baker, H., Zou, J., Snyder, M. P., Gladyshev, V. N., Wyss-Coray, T., Ying, K.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Human health is not a single snapshot that a doctor can take with a single test. It is a continuous, shifting story that unfolds over decades, written in the language of blood chemistry, organ function, genetic code, and daily habits. A blood test captures a moment in time, an MRI scan reveals the structure of an organ, and a diagnosis marks a specific point where a disease becomes visible. Yet, the underlying state of a person's well-being exists between these moments, changing continuously even when no one is looking. The challenge for modern medicine has been to find a way to track this invisible, evolving state, to understand how the body moves from health to illness, and to predict where that path might lead before a crisis occurs.

Researchers have long known that the body's systems are deeply interconnected. A problem in the heart often echoes in the kidneys; a genetic tendency might only reveal itself when combined with a specific lifestyle. Traditional medical models often look at these pieces in isolation or rely on static risk scores that do not update as new information arrives. They struggle to connect the dots between a blood test taken today, an MRI from five years ago, and a new diagnosis tomorrow. The question remains: can we build a system that learns the hidden rules of human biology, integrating every fragment of data into a single, living picture of a person's health that updates itself as new evidence comes in?

A team of scientists has answered this question with a new approach called HealthFlux. They built a "world model," a type of artificial intelligence that does not just memorize past diseases but learns the underlying dynamics of how health changes over time. To train this system, they fed it a massive amount of information from more than 500,000 people in the UK Biobank. This data included nearly 5,000 different types of measurements, ranging from routine blood tests and lifestyle habits to complex three-dimensional images of the brain and body, and genetic profiles. The model learned to weave these disparate threads together into a single, continuous representation of each person's health state.

Unlike standard medical tools that wait for a new test result to update their view, this system simulates what happens in the time between tests. It uses a mathematical engine to project how a person's health state evolves day by day, based on their age and permanent traits like genetics. When a new piece of data arrives, such as a new diagnosis or a blood test result, the system instantly revises its understanding, much like a doctor updating a patient's file with new findings. This allows the model to maintain a consistent, up-to-date picture of a person's health even during long periods without medical contact.

The results of this approach are striking. When the researchers tested the model on people it had never seen before, it successfully predicted the future risk of 1,256 different diseases and death over a twenty-year period. In these tests, the model outperformed existing state-of-the-art tools, including specialized risk scores for heart disease and general machine learning models. For example, among the people the model identified as having the highest risk for heart failure, nearly 30% were diagnosed with the condition within five years, compared to less than 2% of the general population. Similarly, for type 2 diabetes, the model correctly identified the vast majority of future cases years before they were officially diagnosed.

One of the most significant findings is that the model learned the concept of health itself, rather than just memorizing specific diseases. The researchers tested this by hiding certain diseases from the model during its training phase. Even without ever seeing the labels for these specific conditions, the model could still predict who would develop them, ranking future patients correctly in 78 out of 100 comparisons. This suggests the system has learned a fundamental map of human biology that applies across different conditions, allowing it to recognize patterns of decline that signal trouble, even for rare or newly defined illnesses.

The model also proved capable of simulating the effects of medication. By treating the start of a drug as an action that alters the health trajectory, the researchers could run virtual trials to see how different treatments might change a person's future. They simulated the effects of common drugs like statins and blood pressure medications and found that the predicted changes in biological markers, such as cholesterol levels, matched the results reported in hundreds of real-world randomized clinical trials. Furthermore, the model predicted that certain medications could reduce the risk of a wide range of diseases, not just the ones they were originally prescribed for, hinting at new possibilities for drug repurposing.

Crucially, the study showed that combining different types of data is essential. The model identified future cases of hypertension that were missed by models looking at only one type of data, such as blood pressure readings alone. By integrating genetic information, lifestyle factors, and imaging, the system found hidden risks that single-modality approaches overlooked. The researchers also demonstrated that the model's predictions align with the physical reality of the body; for instance, the brain regions the model flagged as important for predicting heart disease were the same regions that showed structural differences in people who later developed the condition.

This work represents a shift from viewing health as a series of disconnected events to understanding it as a continuous, evolving process. By creating a digital twin of human health that learns from the entire spectrum of biological data, the researchers have built a tool that can see further into the future than ever before. While the model is not a crystal ball and its predictions are based on statistical patterns rather than guaranteed outcomes, it offers a powerful new way to identify individuals at risk long before symptoms appear. The ability to simulate the future of a patient's health, to test the impact of interventions in a virtual environment, and to spot the early signs of disease across the entire body, opens a new chapter in preventive medicine, moving the focus from treating sickness to preserving health.

Technical Summary: HealthFlux – A Pan-Modal World Model for Human Health

Problem Statement

Human health is a latent, underlying state that cannot be observed directly. Clinical records, biological measurements, and imaging capture only partial, asynchronous, and heterogeneous facets of this state over time. Existing approaches often treat disease prediction as a static classification task or rely on specific modalities (e.g., EHR sequences or single-omics), failing to integrate diverse data sources into a unified representation that persists between observations and evolves continuously. The challenge lies in inferring a latent health state that integrates multimodal evidence, models the continuous dynamics of physiology between irregular measurements, and supports robust prediction of disease and mortality even when specific disease labels are scarce or absent.

Methodology

Data and Scope

The model, named HealthFlux, was trained on data from 502,166 participants in the UK Biobank. It integrates 11 data domains:

Longitudinal clinical records (EHR events).
Demographics, genetics (principal components), and lifestyle.
Blood biochemistry, haematology, proteomics, and metabolomics.
Physical function measurements.
Three-dimensional brain and body MRI volumes.

The input consists of 4,877 time-dependent features (clinical events and numeric values), 43 permanent context variables, and volumetric MRI data.

Architecture: A Hybrid State-Space World Model

HealthFlux employs a continuous-time recurrent state-space architecture designed to mimic clinical reasoning:

Continuous Evolution (Neural ODE): Between observations, the deterministic component of the health state evolves via a Neural Ordinary Differential Equation (ODE) conditioned on age and permanent context. This allows the model to simulate physiological trajectories during intervals without clinical contact.
Observation-Driven Revision (Stochastic Gate & GRU): When new data arrives (e.g., a blood test, diagnosis, or MRI), a stochastic gate weighs the new evidence against the expected state. A Gated Recurrent Unit (GRU) then updates the deterministic state, effectively revising the health assessment based on new findings.
Latent State Representation: The final health state is a 384-dimensional vector formed by concatenating a 256-dimensional deterministic component and a 128-dimensional stochastic component.
Action Branch (Medication Simulation): A separate branch, trained with the world model frozen, represents medications as actions. It uses drug features (ATC codes, molecular targets, chemical structures, and text embeddings) to shift the health state, enabling the simulation of "medication-on" versus "medication-off" trajectories for the same individual.

Training Objectives

World Model: Trained to predict the identity and timing of the next clinical event (next-event cross-entropy and time prediction) using a variational objective. It does not use disease labels for the core state learning.
Readouts: Disease risk (1,256 diseases and death) and a Health State Index (HSI) are predicted via decoders trained on the frozen latent state.
Medication Effects: Biomarker changes and disease risks are simulated by comparing trajectories with and without medication actions, calibrated against randomized controlled trial (RCT) meta-analyses for specific pathways (e.g., LDL, blood pressure).

Key Contributions

Pan-Modal Integration: HealthFlux is the first model to jointly learn from EHR, multi-omics, lifestyle, and 3D MRI volumes within a single continuous-time latent state, demonstrating that integrating modalities identifies at-risk individuals missed by single-modality models.
Generalization to Unseen Diseases: The model can predict risks for diseases excluded entirely from training (both from the world model and the readout). It achieves an AUROC of 0.783 for "strictly unseen" diseases, ranking future cases above non-cases in 78% of pairs, suggesting it learns the underlying dynamics of health rather than memorizing disease labels.
Virtual Clinical Trials: By conditioning state transitions on medication actions, HealthFlux simulates physiological and disease-risk changes. Simulated biomarker changes correlated with observed changes in 286 RCTs ( $r_w = 0.31$ ), and predicted disease risk directions matched observed trial outcomes in 228 RCTs ( $r_w = 0.32$ ).
Long-Horizon Prediction: The model maintains predictive performance over long horizons without new data. In a 20-year simulation without further observations, it outperformed state-of-the-art models (e.g., Delphi-2M), ranking a future case above a non-case in 84% of pairs compared to 58% for the comparator.

Key Results

Disease Prediction: In held-out UK Biobank participants, HealthFlux achieved a mean 5-year AUROC of 0.832 and average precision of 0.032 across 1,256 diseases and death, outperforming specialized clinical scores (QRISK3, SCORE2) and other machine learning models (XGBoost, LightGBM, MILTON, Delphi-2M, RisQ).
Risk Stratification: For the median disease, the top 10% of predicted risk accounted for 51% of new cases over five years. For heart failure, the top 1% had a 29% incidence rate within five years (vs. 1.4% overall).
Health State Index (HSI): A single scalar score (0–1) derived from the latent state stratified mortality and multimorbidity. A 1-SD lower HSI was associated with a 1.67-fold higher hazard of death. The HSI was lower in future cases of chronic conditions up to 10 years before diagnosis.
External Validation: Performance was comparable or superior in four independent US cohorts (CARDIA, ARIC, WHI, All of Us), even in zero-shot settings (All of Us) where the model was applied without retraining.
Modality Attribution: Ablation studies confirmed that each modality contributes unique information. For example, proteomics and metabolomics were critical for metabolic diseases, while physical function and lifestyle were key for mental and circulatory disorders.
Medication Screening: The model identified candidates for drug repurposing, such as metformin and bisoprolol, predicting broad reductions in disease risk across hundreds of conditions.

Significance and Claims

The authors claim that HealthFlux treats health itself as the object of prediction. By learning the latent dynamics of health from asynchronous, pan-modal data, the model creates a persistent, revisable state that supports risk estimation years before diagnosis.

The paper emphasizes that:

The model learns generalizable health dynamics rather than specific disease patterns, evidenced by its ability to predict diseases it was never trained on.
Integration is essential: Single-modality models miss a significant fraction of future cases (28–56% for hypertension) that the pan-modal model captures.
Simulation is feasible: The model can simulate the physiological effects of medications and predict downstream disease risks, aligning with evidence from randomized trials.
Clinical Utility: The HSI provides a unified metric for health status that correlates with self-rated health and predicts mortality and multimorbidity, potentially enabling earlier intervention.

The authors note that these are retrospective predictions based on observational data and that the medication effects are simulations requiring validation in prospective clinical studies. The model is presented as a tool for generating hypotheses and identifying high-risk individuals, not as a replacement for clinical judgment or a certified diagnostic device.