Causal Workflow AI: Learning Clinical Care Pathways for Safe and Trustworthy Decision Support

Artificial intelligence has made impressive progress in healthcare, yet real-world clinical adoption remains limited due to a fundamental disconnect between static prediction models and dynamic clinical workflows. Current systems fail to capture the sequential nature of medical decision making, lack causal understanding of treatment effects, and provide recommendations that often violate clinical safety constraints. This leads to clinician mistrust and limited practical utility. We introduce CausalCare, a comprehensive framework that bridges this gap through integrated causal inference, temporal modeling, and explicit safety validation. Our approach learns clinical care sequences from multimodal EHR data by constructing causal workflow graphs that respect temporal precedence, incorporate domain knowledge, and enforce safety constraints through a multi-layered validation system. CausalCare provides transparent, step-level explanations aligned with clinical reasoning patterns through three complementary mechanisms: causal pathway visualization, safety rationale presentation, and temporal context analysis. Extensive validation across four large-scale clinical datasets (MIMIC-IV, eICU, OMOP-CDM, MIMIC-CXR) demonstrates superior performance in predicting clinically appropriate next actions (F1-score: 0.81 vs 0.76 best baseline) while reducing unsafe recommendations by 69% compared to state-of-the-art baselines. Our framework achieves particular strength in complex scenarios requiring causal understanding, such as medication sequencing and diagnostic test ordering. A comprehensive ablation study confirms the synergistic contributions of causal learning, safety constraints, and temporal modeling, with the integrated framework outperforming individual components by 11-16%. Through detailed case studies in sepsis management, heart failure, and depression treatment, we demonstrate CausalCare’s ability to respect clinical sequencing, avoid contraindications, and provide interpretable decision support. The framework establishes a new paradigm for trustworthy clinical AI that aligns with human reasoning, workflow patterns, and safety imperatives, representing a significant advancement toward clinically deployable decision support systems.