Explainable AI Analysis of a Prediction Model for Detecting Premature Atrial and Ventricular Complexes

The relationship between premature atrial complexes (PACs) and cardiovascular diseases remains elusive, with existing PAC detectors based on beat classification demonstrating low sensitivity. PAC detectors based on Machine Learning (ML) often lack interpretability, impeding their adoption by cardiologists. Using Explainable AI (XAI) techniques, such as SHapley Additive exPlanations (SHAP), this study enhances the interpretability of a highly accurate (avg. accuracy: 0.985) PAC detector. This detector utilizes a random forest (RF) classifier for normal (N), supraventricular (S), and ventricular (V) beats using ECG-derived features, including heart rate variability (HRV) and QRS complex morphology. Our findings reveal that RR interval features predominantly influence the detection of N and S classes, while QRS morphology critically impacts V class predictions. By refining the model to use only 14 key features from an original set of 185, we developed simpler, surrogate models by using RF and decision trees. Although there is a slight decline in performance-most notably in the sensitivity and positive predictive value (PPV) for S class detection-these models maintain substantial predictive power, hence, underscoring the potential of XAI in building interpretable PAC detectors.