HealthFormer Paper: Generative Multimodal Model Simulates Interventions, Outperforms Clinical Scores

anshulkundaje · x · 2026-08-22

Overview

Eran Segal et al. released a preprint on HealthFormer, a generative multimodal transformer designed to model human physiological trajectories.

Model & Data

Architecture: Decoder-only transformer.
Data: Trained on the deeply phenotyped Human Phenotype Project cohort (>15,000 individuals).
Modalities: 667 measurements across 7 domains (blood biomarkers, body comp, sleep, CGM, microbiome, wearables, meds/behavior).
Objective: Single objective—predict the next measurement.

Key Findings

RCT Validation: Matched direction of effect in 41/41 published RCTs; 30/41 within reported 95% CI.
Reconstruction & Forecasting: Reconstructs biomarkers at r > 0.9; forecasts 2 years ahead.
Generalization: Validated on external cohorts including UK Biobank and NHANES.
Benchmark: Outperformed Framingham CVD & PREVENT-ASCVD on 27/30 endpoints.
Intervention Simulation: Capable of in silico simulation of interventions (e.g., personalized nutrition).

Significance

Demonstrates that a single generative objective enables transfer to diverse clinical tasks without task-specific training and allows for intervention simulation.

Original post →

HealthFormer Paper: Generative Multimodal Model Simulates Interventions, Outperforms Clinical Scores

More from Multimodal