Simulating clinical interventions with a generative multimodal model of human physiology
Guy Lutsker, Gal Sapir, Jordi Merino, Smadar Shilo, Anastasia Godneva, Eli Meirom, Shie Mannor, Hagai Rossman
Abstract
Understanding how human health changes over time, and why responses to interventions vary between individuals, remains a central challenge in medicine. Here we present HealthFormer, a decoder-only transformer that models the human physiological trajectory generatively, by training on data from the Human Phenotype Project, a multi-visit cohort of over 15,000 deeply phenotyped individuals. We tokenise each participant's health trajectory across 667 measurements spanning seven domains: blood biomarkers, body composition, sleep physiology, continuous glucose monitoring, gut microbiome, wearable-derived physiology, and behaviour and medication exposure. We train HealthFormer to forecast individual physiological trajectories across these domains, and from this single generative objective a range of clinically relevant tasks can be expressed as queries on the model. We show that, without task-specific training, HealthFormer transfers to four independent cohorts and improves prediction for 27 of 30 incident-disease and mortality endpoints, exceeding established clinical risk scores in every comparison. We further show that the model can simulate interventions in silico: in a held-out personalised-nutrition trial, intervention-conditioned predictions recover individual six-month biomarker changes (e.g., Pearson r = 0.78 for diastolic blood pressure). Across 41 randomised intervention-outcome comparisons drawn from published trials, our results show that the predicted direction of effect agrees in every case, and the predicted mean falls within the reported 95% confidence interval in 30 cases. We position HealthFormer as an initial health world model, from which forecasting, risk stratification, and intervention-conditioned simulation arise as queries, providing a basis for clinical digital twins.
AI Impact Assessments
(3 models)Scientific Impact Assessment: HealthFormer
1. Core Contribution
HealthFormer represents a conceptual shift from modeling electronic health records (discrete clinical events) to modeling continuous physiological trajectories across multiple biological systems. The key novelty is threefold: (a) a unified tokenization scheme that encodes 667 heterogeneous health measurements across seven domains into a single autoregressive sequence; (b) a decoder-only transformer trained with next-token prediction on ~64M tokens from >15,000 deeply phenotyped individuals; and (c) the demonstration that a single generative objective, without task-specific fine-tuning, enables forecasting, risk stratification, and intervention-conditioned simulation as downstream queries.
The most striking claim is that the model can simulate clinical interventions in silico. By appending treatment tokens to a participant's sequence, the model generates predicted biomarker trajectories that agree in direction with 41/41 published RCT comparisons and fall within the reported 95% CI in 30/41 cases. This positions the work at the intersection of digital twins, foundation models, and precision medicine.
2. Methodological Rigor
Strengths: The evaluation strategy is multi-layered and well-structured. The authors evaluate within-visit reconstruction, longitudinal prediction over ~2 years, zero-shot transfer to four independent cohorts (UK Biobank, NHANES, Framingham, PNP3), disease risk prediction across 30 endpoints, individual-level intervention prediction in a held-out nutrition trial, and population-level comparison against 41 RCT endpoints. The progressive context experiment (Fig. 4d) showing monotonic improvement with additional data is convincing. Negative control experiments (Supplementary Fig. S7) demonstrating pharmacological specificity strengthen the intervention claims.
Concerns: Several methodological issues warrant scrutiny. First, the synthetic population approach for RCT validation introduces a significant degree of freedom — synthetic cohorts are generated from published Table 1 statistics, not actual trial participants, making it impossible to assess whether the model captures individual-level heterogeneity versus population-level regression patterns. Second, the PNP3 "held-out" trial shares ascertainment criteria, CGM devices, and diet-logging conventions with HPP, weakening claims of true external validation. Third, the concordance metric (predicted mean within published 95% CI) is generous — many CIs in large trials are wide, and the predicted direction matching 41/41 comparisons could partly reflect learning that medications generally improve the conditions they treat. Fourth, the model architecture is relatively modest (139M parameters, 2 attention heads) which, while practically advantageous, raises questions about whether the scaling analysis truly supports the claim that larger models would improve substantially. Fifth, the intervention encoding is crude — a single categorical token per medication class without dose, and temporal spacing encoding frequency — limiting the clinical fidelity of simulated interventions.
3. Potential Impact
The practical implications are substantial. If the intervention-conditioned prediction framework proves reliable, it could accelerate hypothesis generation for clinical trials, enable personalized treatment selection, and support regulatory science through in silico evidence. The disease risk prediction results (improving on established clinical scores like Framingham CVD and PREVENT-ASCVD across all comparisons) suggest immediate utility for risk stratification in clinical settings.
However, the authors are appropriately cautious: the model learns associations, not causal effects. The gap between "the predicted direction matches RCT results" and "the model can reliably simulate what would happen to a specific patient under treatment X" is enormous. The systematic under-prediction of high-potency interventions (high-dose statins, semaglutide) and over-prediction of off-target effects (SGLT2i on blood pressure) reveals fundamental limitations tied to the observational training distribution.
4. Timeliness & Relevance
This work arrives at a critical inflection point where foundation models are being applied to health data (BEHRT, Foresight, recent Nature papers on generative EHR models) but have been limited to administrative medical records. The availability of deeply phenotyped cohorts like HPP creates a new data substrate. The clinical digital twin concept has attracted enormous interest but lacked concrete computational instantiations. HealthFormer provides perhaps the most complete prototype to date, though the authors wisely frame it as an "initial health world model" rather than a clinical digital twin.
5. Strengths & Limitations
Key Strengths:
Notable Weaknesses:
Missing Analyses:
Summary
HealthFormer is an ambitious and well-executed study that advances the field by demonstrating that generative modeling of dense physiological trajectories enables a richer class of clinical queries than EHR-based approaches. The intervention simulation results, while not establishing causality, represent a meaningful proof-of-concept. The work's impact will ultimately depend on prospective validation and whether the gap between observational learning and causal intervention effects can be bridged through the extensions the authors propose.
Generated May 1, 2026
Comparison History (145)
Paper 1 likely has higher scientific impact due to its novel, large-scale multimodal “health world model” framing, strong methodological evidence (15k longitudinal cohort, 667 measurements, cross-cohort transfer, broad endpoint improvements, and intervention simulation validated against RCTs), and immediate high-stakes real-world applicability (risk prediction, digital twins, in silico trialing). Its breadth spans clinical medicine, systems biology, epidemiology, and ML. Paper 2 is rigorous and timely for agent evaluation, but its impact is narrower (benchmarking/diagnostics for LLM agents in a specific POMDP supply-chain setting) and less directly consequential outside AI evaluation.
While Paper 1 offers valuable theoretical improvements to AI alignment, Paper 2 presents a transformative foundation model for human physiology with profound applications in medicine. By successfully predicting disease endpoints across independent cohorts and accurately simulating clinical interventions in silico, HealthFormer demonstrates extraordinary methodological rigor and massive real-world utility. Its potential to revolutionize personalized medicine, clinical trials, and epidemiological risk stratification gives it a significantly broader and more profound scientific impact across multiple disciplines.
Paper 2 likely has higher scientific impact: it introduces a large-scale generative multimodal “digital twin” model trained on 15k longitudinal, deeply phenotyped individuals, demonstrates transfer to independent cohorts, improves many clinical endpoints vs established risk scores, and shows intervention-conditioned simulation aligned with RCTs. The real-world applications (risk stratification, forecasting, in-silico intervention testing) are immediate and broad across medicine and ML. Paper 1 is novel and timely for agent research, but is more conceptual and task-family specific, with narrower near-term applicability.
HealthFormer represents a paradigm-shifting contribution to precision medicine by creating a generative foundation model trained on deeply phenotyped longitudinal data from 15,000+ individuals across 667 measurements. Its ability to simulate clinical interventions in silico, validated against 41 published RCTs, and transfer to four independent cohorts without task-specific training, has enormous potential for clinical digital twins, drug development, and personalized medicine. The breadth of impact across healthcare, the scale of validation, and the transformative clinical applications substantially exceed Paper 2's contributions, which, while innovative in autonomous scientific discovery, are demonstrated on a narrower optical physics platform with more incremental results.
While Paper 1 presents a brilliant algorithmic advance for chaotic dynamical systems, Paper 2 demonstrates immense, immediate potential to revolutionize medicine. By creating a multimodal foundational model capable of forecasting disease risk and simulating clinical interventions in silico, it addresses critical bottlenecks in personalized medicine and clinical trial design. Its rigorous validation across multiple independent cohorts and 41 randomized trials suggests unprecedented real-world applicability in healthcare, granting it vastly broader societal and scientific impact.
HealthFormer represents a fundamentally novel approach to modeling human physiology as a generative process, enabling clinical digital twins and in silico intervention simulation. Its validation across independent cohorts, 30 disease endpoints, and 41 randomized trial comparisons demonstrates extraordinary breadth and rigor. The potential to personalize medicine by simulating interventions before administering them has transformative real-world clinical applications. While Paper 1 makes solid engineering contributions to agentic AI training data curation, it is incremental (3.9% improvement) and narrower in scope. Paper 2 opens an entirely new paradigm in precision medicine with cross-disciplinary impact spanning AI, clinical medicine, and public health.
Paper 1 likely has higher scientific impact due to its novel “health world model” framing and strong real-world translational potential (risk prediction and intervention simulation/digital twins) backed by large multimodal longitudinal data, multi-cohort transfer, and comparisons to clinical risk scores plus trial-aligned intervention effects. Its breadth spans clinical medicine, epidemiology, and multimodal ML, and is timely for precision health. Paper 2 is innovative and relevant for agent RL, but its impact is more bounded to GUI agent training and depends heavily on evaluator reliability and benchmark generality.
HealthFormer represents a paradigm-shifting contribution to precision medicine by introducing a generative foundation model for human physiology trained on deeply phenotyped longitudinal data. Its ability to simulate clinical interventions in silico, validated against 41 randomized trial comparisons, and its transfer to independent cohorts for disease/mortality prediction, positions it as a foundational tool for clinical digital twins. The breadth of impact across medicine, clinical trials, and personalized health is enormous. SPADE, while methodologically elegant and useful for scientific ML, addresses a narrower technical problem (structure-prior enforcement) with more limited real-world applicability.
Paper 1 likely has higher scientific impact: it proposes a large-scale generative “health world model” trained on longitudinal multimodal human physiology with demonstrated cross-cohort generalization, improved incident disease/mortality prediction over clinical risk scores, and credible in silico intervention simulation aligned with RCT effects—clear pathways to major real-world clinical applications (risk stratification, digital twins, trial planning). Methodologically it leverages rich multi-domain data and external validations. Paper 2 is novel and timely for efficient reasoning in LLMs, but its impact is primarily within ML systems/efficiency and less broadly transformative than potential clinical translation.
Paper 2 likely has higher scientific impact due to a novel, broadly applicable “health world model” trained on large longitudinal multimodal data, with strong cross-cohort transfer and clinically relevant evaluations (risk prediction, incident endpoints, and intervention-conditioned simulation). Its real-world applications (forecasting, risk stratification, digital twins, trial emulation) are immediate and high-stakes, spanning medicine, public health, and ML. Paper 1 is timely and important for methodology in LLM assessment, but its impact is narrower (measurement validity for psychometrics on LLMs) and less directly translational.