A Causal Explanation Framework for Clinical AI: Methods and Technical Evaluation

Objective: Current AI explanation methods in clinical decision support-such as semantic similarity and feature attribution-provide associative justifications but do not explicitly represent the causal reasoning structures that clinicians use in diagnostic practice. We developed and evaluated a framework for generating structured, evidence-linked causal explanations in clinical AI. Materials and Methods: Using MIMIC-III data and standardized medical ontologies (SNOMED CT, UMLS), we developed a framework that maps AI predictions to domainspecific causal graphs, assembles temporally ordered causal pathways, and links relationships to clinical guidelines and literature. We evaluated framework behavior across 100 acute myocardial infarction (AMI) cases, comparing two instantiations (domain-specific causal graph and large-scale medical knowledge graph [LMKG]-enhanced) with commonly used explanation approaches (BioBERT semantic similarity and SHAP feature attribution). Evaluation focused on explanation expressiveness and structural properties. Metric robustness was assessed using adversarial degradation tests. Results: The framework consistently generated structured causal explanations across all evaluated cases, achieving high scores on causal reasoning-oriented metrics (fidelity 0.85±0.01; completeness 1.00±0.00). LMKG integration expanded identifiable causal pathways while preserving core explanation structure. In contrast, similarity and attribution methods were unable to express mechanistic causal pathways by design, reflecting their complementary explanatory purposes. Adversarial validation demonstrated that metrics discriminated between intact and degraded explanations (mean AUC-ROC 0.66). Independent case validation revealed that structural reliability depends critically on knowledge base completeness. Discussion: These results demonstrate the technical feasibility of representing causal reasoning structures within clinical AI explanations using ontologies and evidence-linked causal graphs. The framework provides capabilities complementary to existing explanation approaches. Conclusion: This work demonstrates the technical feasibility of causal explainability as a distinct dimension of clinical AI interpretation. The framework generates structured causal pathways grounded in clinical knowledge, providing a template for evidence-based explanation systems. Important: This is a technical proof-of-concept, not a validated clinical tool. Prospective clinician validation is required to assess clinical utility and impact on decision-making before any deployment.