Explainable AI Identifies and Localizes LV Scar in HCM Using 12-Lead ECG

Authors, Journal, Affiliations, Type, DOI

Overview

LV scar (detected by LGE-MRI) is a major risk factor for sudden death and heart failure in HCM, but MRI is expensive, unavailable worldwide, and artifact-prone with implanted devices. This paper introduces XplainScar, an explainable ML framework that detects and localizes LV scar from standard 12-lead ECG. Trained on 500 HCM patients from Johns Hopkins and externally validated on 248 from UCSF, XplainScar achieves F1=89%, sensitivity=91%, specificity=78% on the held-out set, running in <1 min for 10 patients on a standard PC. The model uses unsupervised ECG clustering (Dirichlet Process GMM), self-supervised contrastive learning (SCARF), and SHAP-based explainability to identify regional ECG signatures of basal, mid, and apical LV scar.

Keywords

Left ventricular scar, Hypertrophic cardiomyopathy, Machine learning, Explainable AI, Unsupervised clustering, Self-supervised learning

Key Takeaways

Background and Clinical Motivation

Patient Population

XplainScar Pipeline (5 steps)

  1. ECG feature extraction: 23 features per lead (QRS duration, amplitude, energy, AUC; ST segment; T wave amplitude, inversion, energy; TP segment slope) → 276-dimensional feature vector per patient
  2. Confounder adjustment: Multiple linear regression to remove LVMI, age, sex effects from ECG features during training only (not applied at inference)
  3. Unsupervised clustering: Dirichlet Process GMM partitions patients into homogeneous ECG-similarity subgroups; recursive merging until 2 merged groups per regional scar task; this step critical — removing it drops F1 from 92% to 41%
  4. Self-supervised + supervised learning (SCARF): Per-cluster neural network using SCARF (corrupts 60% of features randomly, contrastive pretraining), then fine-tunes with cross-entropy classification; self-supervised step adds ~10% F1
  5. Explainability (SHAP): Shapley values assigned to each of the 276 ECG features per prediction, identifying which features pushed toward Scar vs NoScar

Model Performance

ECG Signatures of LV Scar (SHAP Analysis)

Scar Burden and Detection Sensitivity

Cross-Modal Validation

Longitudinal Testing

Limitations of the Document

Key Concepts Mentioned

Key Entities Mentioned

Wiki Pages Updated