AI Characterisation of Discordance Profiles Between Stress Electrocardiogram and Myocardial Tomoscintigraphy Using Random Forest XGBoost and SHAP

Purpose: The evaluation of Coronary Artery Disease (CAD) using stress ECG and Myocardial Perfusion SPECT (MPS) frequently reveals discrepancies, particularly in patients with a positive ECG and a negative MPS (MPS-/ECG+). This paradoxical group poses a diagnostic challenge, often unexplained by classic statistical analyses. Our study aims to characterize this profile using interpretable Artificial Intelligence (AI). Patients and Methods: A cohort of 86 patients with negative MPS was stratified into a "discordant" group (MPS-/ECG+, n=19) and a "concordant" group (MPS-/ECG-, n=67). A two-pronged analytical approach was used: (1) bivariate statistical analysis and (2) machine learning modelling. Two algorithms, Random Forest and XGBoost, were trained on a dataset rebalanced using SMOTE. Performance was assessed by AUC, and interpretability was ensured by SHAP analysis. Results: Conventional univariate statistical analysis did not identify any variable significantly associated with discordance (all p>0.05). In contrast, the XGBoost model, exploiting multivariate interactions, surpassed Random Forest, achieving a moderate but informative performance (AUC = 0.78), with a Sensitivity (Recall) of 67%, a Precision (Positive Predictive Value) of 50% for the discordant class (Class 1), and an overall Accuracy of 76%. SHAP analysis revealed that the most important predictors of discordance were female gender, advanced age, and the presence of diabetes, indicating that the prediction was influenced by the combination of these factors rather than a single one. Conclusion: This exploratory study demonstrates the potential value of explainable AI for deciphering complex clinical problems such as MPS/ECG discordance. Our multivariate models identified a potential patient profile associated with of MPS-/ECG+ discordance, characterized by the synergy of clinical factors. These preliminary results suggest that ECG abnormalities in the absence of a perfusion deficit might reflect an underlying pathology (e.g. microvascular disease) rather than a simple false positive, warranting further prospective validation.