- The paper presents HealthFormer, a generative multimodal transformer that simulates individual physiological responses to clinical interventions.
- It uses an autoregressive framework with tokenized longitudinal data across 667 modalities to achieve robust, cross-system forecasting.
- The model accurately simulates intervention outcomes and risk stratification, paving the way for digital twin applications in precision medicine.
Simulating Clinical Interventions with a Generative Multimodal Model of Human Physiology
Introduction and Motivation
The simulation of longitudinal physiological trajectories and intervention responses at the individual level is a central but unresolved challenge in precision medicine, where conventional models are often constrained by episodic and modality-specific data, limited direct physiological measurement, and an inability to model dynamic cross-system responses. HealthFormer addresses these deficiencies by constructing a generative, sequence-level transformer model trained on a uniquely broad and deeply phenotyped multimodal longitudinal dataset, yielding a unified predictive framework for human physiological dynamics and intervention simulation.
Data Sources and Model Architecture
HealthFormer is pre-trained on the Human Phenotype Project (HPP), comprising over 15,000 participants with repeated, protocolized measurements across 667 modalities spanning blood biomarkers, body composition (including DEXA), sleep physiology, continuous glucose monitoring, gut microbiome, wearable sensor data, behavioral, and medication exposures. Each participant’s full longitudinal health history is tokenized as a unified, temporally ordered sequence, facilitating autoregressive next-token prediction with a vocabulary of over 13,000 tokens.
The architecture is a decoder-only transformer with 139 million parameters, a 25,000-token sequence length, and a hierarchical embedding scheme encompassing token, modality, temporal, positional, continuous-value, and demographic features. The model employs a next-token objective, with specialized loss functions for continuous and categorical modalities, augmented for longitudinal forecasting via a split-context loss and robustified with stochastic data augmentations. A key methodological advance is the explicit handling of the heterogeneity, sparsity, and temporality of multimodal clinical data without imputation, leveraging a tokenization and attention strategy that preserves both content and temporal structure, as well as enabling targeted queries for future or hypothetical outcome prediction.
HealthFormer’s unified representation enables autoregressive reconstruction of observed health state within visits (within-visit: category-level mean Pearson r up to 0.82 for ABI; individual-biomarker within-visit r > 0.9 for multiple clinical measures), longitudinal forecasting of future physiological states across a two-year horizon (longitudinal: mean r ≈ 0.65–0.81 across most stable modalities, lower with increasing temporal drift or measurement sparsity), and robust recovery of cross-system associations (population-level distributions r up to 0.97 for e.g., pelvic area to body weight or BMI to brachial pressure). Notably, longitudinal forecasting with HealthFormer exceeds independently trained per-modality regression models in 89.5% of modalities with significant differences, demonstrating the utility of shared physiological structure and cross-modal integration.
Performance scales with model size—mean r in longitudinal prediction improves monotonicly with increased model capacity, supporting further gains from larger architectures and data. The model’s cross-modal conditional inference is empirically validated, supporting both expected and non-obvious physiological dependencies without explicit pairing during training.
Generalization and Clinical Risk Prediction
Without task-specific adaptation, HealthFormer generalizes to external population cohorts (UK Biobank, NHANES, PNP3, Framingham), achieving substantial predictive skill in shared modalities—median zero-shot Pearson r = 0.52 (UKB), 0.39 (NHANES), 0.70 (PNP3), 0.45 (Framingham). Downstream risk models leveraging HealthFormer’s embeddings from UK Biobank baseline data improve concordance on 27 of 30 incident disease and mortality endpoints, outperforming both age+sex+BMI baselines and established disease-specific clinical risk scores in all tested comparisons. For instance, concordance indices for major outcomes include C = 0.826 for cardiovascular mortality, C = 0.834 for chronic kidney disease, and C = 0.808 for heart failure.
HealthFormer-derived biological age is a robust stratifier for prevalent metabolic and cardiovascular disease in NHANES, matching or exceeding demographic baselines for 7/8 prevalent disease conditions.
Simulation of Clinical Interventions and Alignment with Randomized Trials
A hallmark capability of HealthFormer is intervention-conditioned simulation—supporting both individual-level and synthetic population-level prediction under hypothetical or observed interventions. In a held-out CGM-guided personalized nutrition RCT (PNP3), HealthFormer’s six-month predictions for BMI, fasting glucose, and diastolic blood pressure track observed individual changes with r = 0.63, 0.61, and 0.78, respectively. Across 10 biomarkers in PNP3, 9/10 show lower mean absolute error than a no-change baseline; predictions improve monotonically with increasing pre-outcome measurement context.
For external validation of intervention simulation, 41 RCT comparisons across lipid-lowering, antihypertensive, glycaemic, weight-loss, exercise, and combination interventions were performed by generating trial-matched synthetic populations and appending intervention tokens. The predicted direction of effect agreed with published endpoints for all 41 comparisons; the predicted mean fell within the published 95% confidence interval for 30/41 (73%). Largest effect-size endpoints (e.g. high-dose statins, GLP-1s) are underpredicted, interpretable as a manifestation of population-level effect size distributions in the observational pretraining data and representation of dosing only at the tokenized class level. The model recovers temporal dose-response relationships, cross-system off-target effects, and active comparator potency ranking for drug classes.
The model also handles negative controls, with no spurious prediction of effect for pharmacologically unrelated intervention–outcome pairs, and distinguishes effect magnitude according to physiological pathway specificity.
Model Limitations and Future Directions
HealthFormer is strictly associational, not a causal estimator; its intervention-conditioned outputs provide prognostic stratification within exposure group but do not support counterfactual efficacy estimation for individual patients. Representation of interventions is currently constrained (e.g., drug dose, adherence, intensity, timing are partially encoded), and granular continuous physiological signals (e.g., raw waveforms, high-frequency telemetry) are summarized for tractability.
Clinically actionable digital twins will require further prospective calibration, granular mediation of exposure tokens, and explicit causal inference overlays. Nevertheless, the demonstrated ability to generalize, synthesize, and simulate trajectories and interventions for unseen populations and novel queries establishes HealthFormer as a substrate for downstream clinical hypothesis generation, policy simulation, and augmentation of prospective study design.
Implications and Prospects
HealthFormer operationalizes the concept of a “health world model” and offers a framework where forecasting, risk stratification, and (importantly) simulation of intervention outcomes are unified downstream queries. This blurs the boundary between phenotyping, forecasting, and simulation, setting the stage for generative digital twins that interact with the clinical decision space. As tokenization, granularity, and cohort scale improve, such models are anticipated to underpin practical systems for individualized treatment planning, safety assessment, and early-phase in silico clinical trials. Moreover, the representational approach adopted by HealthFormer is directly extensible to other data-saturated domains involving high-dimensional, sparse, irregular, multi-modal, and temporally resolved signals.
Conclusion
HealthFormer demonstrates the feasibility and utility of generative multimodal modeling for simulating both physiological trajectories and clinical interventions in human health. Leveraging large-scale, deeply phenotyped, longitudinal data for unified cross-modal autoregressive modeling, it achieves strong predictive skill in both individual and synthetic population settings, across both internal and external validation cohorts. Its capacity for intervention-conditioned simulation, including broad concordance with external randomized clinical trial evidence, positions it as a foundational model for subsequent development of clinical digital twins and individualized, data-driven medicine. The progression toward increasingly granular, causally-calibrated, and prospectively validated world models of human physiology is supported by this work’s methodological rigor, empirical depth, and conceptual advances (2604.27899).
