Diagnosis of Chagas disease using tensor Tucker-compressed ECG scalograms

Abstract Objective: As modest illness-related patterns are dispersed throughout the temporal, spectral, and inter-lead dimensions of the electrocardiogram (ECG) signals, automated diagnosis of Chagas disease from ECG recordings remains difficult. To address this challenge, this study develops a tensor-based representation framework that preserves the multidimensional characteristics of ECG scalograms while learning low-dimensional projection subspaces using Tucker decomposition. Approach: The proposed framework integrates multi-domain feature extraction, tensor-based representation learning, dimensionality reduction, and ensemble classification for automated Chagas disease detection from 12-lead ECG recordings. It extracts five categories of features: (a) scalogram from specific ECG leads;(b) scalogram features from beat-aligned averaged ECG segments; (c) heart rate variability-based statistics;(d) statistical features from every ECG lead; and (e) patient information (weight, age, and sex). The continuous wavelet transform (CWT) for time-frequency feature extraction is used to create scalograms, which allow for the capture of multi-scale frequency patterns and local transients linked to the course of the disease. A 3-D tensor representation that maintains spatial, temporal, and frequency linkages was created by assembling the scalogram-based features. Optimal partial Tucker tensor decomposition was used to reduce dimensionality. For improved classification performance, the resulting core tensor was flattened, merged with the remaining features (b to e), and then subjected to Fisher score-based feature selection. A RUSBoost classifier was employed for binary classification of 12-lead ECG recordings into Chagas and other classes, while addressing class imbalance through random under-sampling. Main results: The proposed framework was evaluated on the hidden PhysioNet/CinC 2025 Challenge test set through a follow-up Challenge submission under the team name Spectra, achieving an accuracy of 88.7%, an F1-Chagas score of 0.171, an AUPRC of 0.136, an AUROC of 0.631, and an overall Challenge Score of 0.217, indicating good generalization to previously unseen ECG recordings. Significance: This work introduces a projection-based tensor representation learning strategy for ECG analysis that transforms high-dimensional scalogram based features into compact latent features while retaining clinically relevant temporal, spectral, and inter-lead information. The proposed framework can facilitate reliable automated Chagas disease screening from routine ECG recordings for earlier clinical intervention.