Causal Foundation Models for Trustworthy Multi-Modal Clinical Decision Intelligence: Integrating Medical Imaging, Electronic Health Records and Explainable Reasoning

Abstract

Modern clinical artificial intelligence (AI) systems overwhelmingly rely on correlational, single-modality pattern recognition, a design choice that undermines robustness, fairness and clinician trust when models are deployed outside their training distribution. This paper proposes a Causal Foundation Model (CFM) framework for multi-modal clinical decision intelligence that jointly reasons over medical imaging and structured electronic health record (EHR) data through an explicit causal structure-learning layer, coupled with post-hoc and structure-grounded explainability. Unlike purely associative deep learning pipelines, the proposed architecture (i) learns a shared multimodal representation through cross-modal attention fusion, (ii) recovers a candidate causal graph over clinical variables using constraint-based causal discovery and (iii) grounds every prediction in both feature-attribution and causal-path rationales that a clinician can audit.
We instantiate and evaluate a proof-of-concept version of the framework on the open UCI Heart Disease dataset (303 patients, 13 clinical variables) as the EHR modality, coupled with a clearly labelled synthetic radiographic-marker channel used only to demonstrate the multimodal fusion mechanism. Causal discovery using the PC algorithm recovers a clinically plausible skeleton in which chest pain type, exercise-induced angina, thalassemia status, ST-depression (oldpeak) and number of major vessels (ca) emerge as the strongest neighbours of the diagnostic target, consistent with established cardiology literature. A causally-informed feature subset achieves 83.8% accuracy and 0.914 AUC-ROC under five-fold stratified cross-validation, marginally outperforming both the unimodal EHR baseline (82.2% accuracy, 0.909 AUC) and the naively fused multimodal model (83.5% accuracy, 0.907 AUC), while using fewer than half the input features — evidence that causal feature selection can match or exceed brute-force fusion at substantially lower complexity.
SHAP-based attribution further shows that clinically established risk indicators dominate the model's decisions, strengthening the case for causal, explainable architectures as a template for trustworthy clinical decision support. We discuss the limitations of the present proof-of-concept, outline a roadmap toward credentialed large-scale multimodal validation using resources such as MIMIC-CXR and MIMIC-IV and argue that causal grounding, rather than scale alone, is the more defensible route to clinically deployable, regulator-facing AI.