Interpretable Multimodal Stratification of Chagas Disease Patients Using RR-Interval Features from Low- and High-Frequency Bands

FRANCO AMARU CRUZ BARRETO1, Santiago Ismael Flores-Chavez2, Carlos Eduardo Salazar Coaguila3, Antonio Gabriel Ravelo-Garcı́a4, Victor Andres Cabrera-Caso5, Miguel Vizcardo2
1UNSA, 2Universidad Nacional de San Agustin de Arequipa, 3Universidad Nacional de San Agustin, 4Institute for Technological Development and Innovation in Communications, Universidad de Las Palmas de Gran Canaria, 35017 Las Palmas de Gran Canaria, Spain Interactive Technologies Institute (ITI/LARSyS and ARDITI), 9020-105 Funchal, Portugal, 5Facultad de Medicina, Universidad Nacional de San Agustín de Arequipa


Abstract

Background Cardiac autonomic dysfunction is central to Chagas disease, and heart rate variability (HRV) offers a low-cost, noninvasive way to quantify this impairment. Prior work has shown that disease progression is reflected in altered short-term HRV, abnormal power-law behavior, and nonlinear RR-interval dynamics, suggesting that combining complementary HRV descriptors is more useful than relying on a single marker. Objective To stratify Chagas disease patients into Control, CH1, and CH2 groups using interpretable features from 24-hour RR-interval recordings, preserving low computational cost and clinical interpretability. Methods 24-hour RR-interval series from 83 controls, 102 CH1, and 107 CH2 patients were processed through a Pan–Tompkins pipeline, segmented into 5-minute windows, and transformed into the frequency domain to derive LF, HF, and LF/HF series. Statistical and geometric descriptors were computed at the patient level, and a decision-tree classifier was trained on 80% of the cohort and evaluated on the remaining 20%. Results and Expected Improvements The baseline model achieved accuracy above 70%, supporting the feasibility of HRV-based screening in Chagas disease. To strengthen these results, future work should incorporate nonlinear descriptors — DFA α1, sample entropy, and Poincaré-plot indices — replace the single train/test split with repeated stratified cross-validation, and apply SHAP-based feature importance to identify the most stable biomarkers. Time-frequency features from wavelet analysis could further capture the nonstationary behavior of 24-hour HRV. Conclusion A simple decision-tree model built from LF/HF-based features achieves meaningful group stratification. The most promising next step is to enrich the model with nonlinear HRV markers and stronger validation, improving discrimination between early and advanced cardiac involvement while preserving interpretability.