AI-Enabled Early Detection of Chemo-Induced Cardiotoxicity Patterns Using ECG Time Series Data: A Simulated Oncology...

Area of Science:

• Cardio-oncology
• Machine Learning
• Biomedical Signal Processing

Background:

• Chemotherapy-induced cardiotoxicity is a significant clinical challenge, often detected late via costly imaging or biomarkers.
• Current surveillance methods are time-consuming, expensive, and identify damage at advanced stages.
• Electrocardiography (ECG) offers a low-cost, non-invasive alternative for early electrophysiological change detection, yet remains underutilized in cardio-oncology.

Purpose of the Study:

• To develop an explainable machine learning (ML) model for early detection of chemotherapy-induced cardiotoxicity.
• To predict cardiotoxicity patterns using single-lead ECG signals.
• To enhance the clinical utility of ECG in cardio-oncology surveillance.

Main Methods:

• A public ECG dataset (n=4997) was preprocessed into 18 features (temporal, morphologic, spectral).
• Random Forest and XGBoost ensemble algorithms were trained and validated.
• Model performance was assessed using ROC-AUC, PR-AUC, and F1-score with bootstrap resampling; interpretability was evaluated via permutation importance and SHAP analysis.

Main Results:

• Both ML models achieved near-perfect classification performance (ROC-AUC, PR-AUC > 0.99; F1-score ≈ 0.986).
• Key predictive features included spectral entropy, band3 frequency, QT surrogate, and peak count.
• These features correlate with early cardiotoxicity indicators such as repolarization instability and autonomic imbalance.

Conclusions:

• An interpretable, feature-driven ML model demonstrates the potential of single-lead ECG for affordable, early detection of chemotherapy-induced cardiotoxicity.
• This approach offers a feasible pathway for precision cardio-oncology, especially in resource-limited or ambulatory settings.
• Single-lead ECG analysis can serve as a clinically relevant tool for proactive cardiac surveillance during chemotherapy.