Heterogeneity in treatment responses to standard treatment regimens poses a critical challenge in clinical practice, compounded by the unobservability of counterfactual outcomes. To address this, we present temporal causal precision medication (TCPM), a generalist framework translating routine physiological indicators into personalized treatment strategies. Central to TCPM is the treatment-free physiological profile (TFPP), a representation of heterogeneous physiological states trained using counterfactual prediction and adversarial learning to model patient’s intrinsic, unbiased physiological state while eliminating the confounding effects of existing therapies from the routine clinical data. Validated across four public datasets and two private datasets, TCPM significantly outperformed standard clinical protocols and state-of-the-art reinforcement learning methods in therapeutic efficacy and physician simulation studies. Furthermore, TCPM identified patient subgroups with divergent response patterns, and elucidated evolving physiological network dynamics to uncover mechanistic heterogeneity. By integrating data-driven precision with actionable physiological insights, TCPM establishes a clinically practical, evidence-based framework for delivering an optimal individual clinical benefit through personalized treatment allocation in acute and chronic diseases, advancing the clinical implementation of precision medication. The Authors develop a generalist framework, Temporal Causal Precision Medication (TCPM), which learns treatment-free physiological profiles from clinical data to estimate unbiased treatment effects. Validated in six cohorts, TCPM improves reveals subtypes driving heterogeneous responses.
Deng et al. (Mon,) studied this question.