Machine learning-based stratification of chagas heart failure severity using ECG power spectral biomarkers.

purposeThis study presents a machine learning methodology to automatically classify heart failure severity in Chagas disease (CD) patients using non-invasive 24-hour ECG-Holter signals.

methodsFollowing American Heart Association (AHA) guidelines, the cohort was stratified into three Left Ventricular Ejection Fraction (LVEF)-based severity groups: Normal (LVEF ≥ 0.50, n=197), Moderate (0.40 ≤ LVEF < 0.50, n=106), and Severe (LVEF < 0.40, n=77), totaling N=380 patients. From short 10-second ECG segments, we extracted eleven spectral features derived from the power spectral density (PSD). Class imbalance was addressed through oversampling applied to the training folds. All classifiers were evaluated over 50 random stratified train-test splits (80/20) across three pairwise tasks (Normal vs. Moderate, Normal vs. Severe, Moderate vs. Severe).

resultsAnalysis revealed a consistent leftward shift in PSD, with increased low-frequency power in more severe cases, consistent with morphological ECG changes including P-wave attenuation, QRS alterations, and ST-segment shifts. Using this spectral biomarker, the best models achieved mean AUC/PR-AUC values of 0.79/0.76 for Normal vs. Severe and 0.83/0.85 for Moderate vs. Severe across 50 random states. The Normal vs. Moderate task showed moderate separability (AUC = 0.75, PR-AUC = 0.72).

conclusionThese findings highlight the potential of power spectral ECG analysis as a low-cost, fully automated tool for risk stratification in CD. The methodology shows promise for improving triage and clinical decision-making in resource-limited settings where CD remains highly prevalent.