ICOM: AN INTERVENTION BASED FOUNDATION MODEL FOR HEALTHCARE

Foundation models for electronic health records are typically pretrained using observational sequence modeling objectives, implicitly entangling patient state with historically realized clinical actions and limiting their ability to support counterfactual reasoning. We introduce the Interventional Clinical Operator Model (ICOM), which reframes EHR pretraining as conditional operator learning over patient trajectories, explicitly modeling how future clinical states change under structured interventions. ICOM employs a dual-stream Transformer architecture to separate patient state and intervention representations and is trained using scalable interventional supervision generated via stochastic graph surgery. Across a range of evaluations, ICOM matches or exceeds strong baselines on standard observational prediction tasks while substantially improving counterfactual prediction accuracy, compositional generalization to unseen intervention combinations, and robustness under population and intervention shift. We further analyze ICOM through the lens of operator learning, providing theoretical insights into identifiability and generalization under observational data. Together, these results demonstrate that incorporating interventional structure at pretraining time yields foundation models that are better aligned with clinical decision-making and more robust to the distributional complexities of real-world healthcare data.