From association to causation: Integrating stable, subgroup-aware SHAP explanations with causal inference for trustworthy AI in coronary bifurcation PCI.

backgroundCoronary bifurcation lesions (CBLs) represent 15-20% of percutaneous coronary interventions (PCIs) and are associated with increased procedural complexity and adverse outcomes. Although machine learning (ML) models show promise in stratifying lesion complexity, their clinical adoption requires not only high accuracy but also robust, equitable, and causally interpretable explanations. Objectives: To develop high-performance ML models for classifying CBL complexity, to rigorously evaluate the stability and subgroup heterogeneity of SHAP-based explanations, and to extend interpretability from associative feature attribution to causal effect estimation using a unified counterfactual framework.

methodsIn a prospective registry of 500 consecutive PCI patients at Chamran and Askariyeh Hospitals, Isfahan, Iran, from March 2023 to January 2025, lesions were stratified into three classes: non-bifurcation (Class 0), simple bifurcation (Class 1), and complex bifurcation (Class 2). Ten key predictors spanning demographics, angiographic morphology, and procedural variables were selected. Four ML models were optimized via cross-validation. SHAP values were used to quantify feature contributions. Interpretability stability was assessed over 1000 bootstrap iterations using the SHAP Stability Index (SSI) and Kendall's τ. Subgroup analyses were stratified by sex, diabetes, and hypertension. A causal inference pipeline, including a directed acyclic graph (DAG), propensity-stratified average treatment effects (ATE), doubly robust augmented inverse-propensity weighting (AIPW), mediation decomposition, and counterfactual simulation, was implemented.

resultsThe Random Forest model achieved near-perfect classification accuracy with high interpretability stability. Side-branch stenosis, heavy calcification, and bifurcation angle were the dominant global predictors. Subgroup analyses revealed clinically meaningful heterogeneity, with diabetic and female patients assigned substantially greater explanatory weight to calcification. Causal analysis confirmed that dual-stent use was associated with increased procedural risk, largely mediated through procedural intensity. Counterfactual simulation demonstrated that reducing side-branch stenosis meaningfully lowered predicted risk.

conclusionIntegrating stable, subgroup-aware SHAP explanations with formal causal inference transforms ML from a predictive tool into a clinically actionable decision-support system. This dual framework enables clinicians to distinguish drivers of treatment selection from true determinants of outcome, supports equitable interpretation across patient subgroups, and quantifies the policy impact of hypothetical interventions, thereby advancing trustworthy, transparent, and individualized care in high-stakes interventional cardiology.