XplainScar: Explainable Artificial Intelligence to Identify and Localize Left Ventricular Scar in Hypertrophic Cardiomyopathy (HCM) Using 12-lead Electrocardiogram

Background/Rationale: Myocardial scar, identified by late gadolinium enhancement (LGE) on MRI, is associated with sudden death in HCM. Unlike ECG, MRI is expensive and adversely affected by artifacts from implanted devices. However, little is known about ECG features of LV-scar in HCM. Objective: Develop an ECG-based explainable machine learning method to identify and localize LV-scar in HCM. Method: We retrospectively studied 500 HCM patients (JH HCM Registry) for model development, and 248 patients (UCSF HCM Registry) for validation. All patients underwent MRI and ECHO within 1 year of ECG. LV-LGE (scar) was assessed using QMass. After excluding RV-insertion-point-LGE, the LV was divided into basal, mid, and apical regions for scar detection. Resting 12-lead ECGs were segmented, features were extracted and adjusted for LV-mass, age, sex. We utilized unsupervised and self-supervised ECG representation learning, where patients are partitioned into groups of several sub-clusters, each sharing similar ECG patterns, but with high separation between scar and no-scar classes. In each group, a self-supervised neural net and a fully connected neural net successfully predicted LV-scar (see Figure) and revealed ECG features of scar. Results: Our method identifies LV-scar in the JH-dataset with high precision (90%), sensitivity (95%), specificity (80%), F1-score (90%), and generalizes well to UCSF-data (precision:88%, sensitivity:90%, specificity:78%, F1-score:89%). The top ECG features for basal-scar are Q-amplitude, Q-slope, non-terminal QRS duration in aVR, and area under QRS and T wave energy in V1-V2. T-wave inversion in V4-V6, area under QRS in V3, TP slope in V3-V4 predicted apical scar. Features selected for mid scar prediction combine features for basal and apical scar. Conclusion: This is the first ECG-based ML model to identify/localize LV-scar in HCM. Our model demonstrates good performance and reveals ECG features of scar in HCM.