Simulating clinical interventions with a generative multimodal model of human physiology
Simulating clinical interventions with a generative multimodal model of human physiology
Authors: Guy Lutsker, Gal Sapir, Jordi Merino, Smadar Shilo, Anastasia Godneva, Eli Meirom, Shie Mannor, Hagai Rossman, Gal Chechik, Eran Segal Date: 2026-04-30 Paper ID: openalex:2604.27899
Summary
HealthFormer is a decoder-only transformer model designed to act as a generative world model for human physiology. By tokenizing longitudinal health trajectories across 667 diverse clinical and lifestyle variables, the model supports forecasting, risk stratification, and intervention-conditioned simulation as unified query tasks. Evaluation on independent clinical cohorts and historical trial data demonstrates its superiority over traditional clinical risk scores and its capacity to simulate individual-level responses to health interventions.
Key Contributions
- Introduced HealthFormer, a decoder-only transformer trained on 667 physiological measurements across seven multimodal domains from 15,000 participants.
- Demonstrated zero-shot transfer capability to four independent cohorts, outperforming clinical risk scores on 27 of 30 incident-disease and mortality endpoints.
- Enabled in silico intervention simulation, accurately predicting individual biomarker changes and aligning with effect directions in 41 clinical trial benchmarks.
Key Concepts
- health-world-model: A transformer-based foundational architecture designed to model longitudinal multimodal health data for forecasting and in silico intervention simulation.
Archivist Review
The paper introduces a compelling framework for generative modeling of longitudinal patient physiology. I approved the ‘Health World Model’ concept as it defines the nascent field of large-scale generative clinical simulation. The model instance itself (HealthFormer) was rejected as being a specific implementation rather than a general methodological concept.
Approved Concepts
- Health World Model: The concept of an architecture acting as a foundational ‘world model’ for complex physiological processes represents a paradigm shift from task-specific medical AI to general-purpose clinical simulation.
Rejected Candidates
- [concept] HealthFormer (
healthformer) - paper_local: This is a specific model implementation rather than a general, reusable concept.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.27899
- paper_id
- 2604.27899
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- health-world-model
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-03T06:03:26Z