Concordance analysis between deep learning model predictions and electrocardiography thresholds from cardiology guidelines

Abstract Deep learning models for the classification of electrocardiograms (ECGs) are able to learn disease-specific patterns, but they are rarely implemented in medical practice due to their "black box" nature. Post-hoc explainable artificial intelligence (XAI) methods compute regions of interest (ROI) which are of importance for a model’s decision making. However, it needs to be further analyzed whether a model focuses on the morphological or rhythmical information within the ROIs. We evaluate a pre-trained ResNet for sinus bradycardia (SB) and sinus tachycardia (ST) classification on the PTBXL dataset using the XAI method Integrated Gradients. We compare the confidence of the model predictions to ECG features used by clinicians using correlation analysis. Correlation is highest for RR intervals (SB: 0.44) and atrial as well as ventricular heart rates (ST: 0.51), with the majority exceeding clinical thresholds for both disorders, indicating that the model learned rhythmical features. Except for QT intervals in ST classification, morphological features such as duration and amplitudes of P-/T-waves do not show any correlation.