Objectives: Chemotherapy-induced cardiotoxicity is still a major clinical problem, usually appearing subclinically before structural or symptomatic cardiac dysfunction appears. Standard surveillance methods use imaging and biomarkers, which are time-intensive and money-intensive and can only identify damage at more advanced levels. Electrocardiography (ECG) provides a low-cost, non-invasive method that can detect early electrophysiological changes but is not fully utilized in cardio-oncology. The present work was designed to build an explainable machine learning model for predicting chemo-like cardiotoxicity patterns at an early stage from single-lead ECG signals.
Methods: A public ECG data set (n=4997 segments) underwent preprocessing and was converted to 18 temporal, morphologic, and spectral features. Two ensemble learning algorithms-Random Forest and XGBoost-were trained and validated with stratified splits. Model performance was assessed with ROC-AUC, PR-AUC, and F1-score with 1000 bootstrap resampling. Feature interpretability was evaluated through permutation importance and SHAP analysis.
Results: Both models scored near-perfect classification (ROC-AUC and PR-AUC>0.99, F1-score ≈ 0.986). Spectral entropy, band3 (high-energy frequency), QT surrogate, and peak count were the top features ranking alongside early cardiotoxicity indicators like repolarization instability and autonomic imbalance.
Conclusions: The feature-driven, interpretable ML architecture suggested here shows that single-lead ECG has the potential to be an affordable and clinically relevant tool for the early detection of chemotherapy-induced cardiotoxicity. The method provides a feasible route toward implementation in precision cardio-oncology, particularly in resource-poor or ambulatory environments.
Keywords: ECG; cardio-oncology; cardiotoxicity; chemotherapy; precision medicine.
Copyright © 2025 Wolters Kluwer Health, Inc. All rights reserved.