EFFICIENT DEEP LEARNING FOR CHAGAS CARDIOMYOPATHYDETECTION: DATA-DRIVEN ECG FEATURE REDUCTION
Keywords:
Chagas cardiomyopathy; Electrocardiogram (ECG); Deep learning; Principal Component Analysis (PCA); Feature reduction; Biomedical signal processing; Medical AIAbstract
Chagas cardiomyopathy, a severe complication of chronic Trypanosoma cruzi infection,
requires early and efficient diagnosis to improve outcomes. While electrocardiograms (ECGs) are
accessible diagnostic tools, their high dimensionality limits the scalability of deep learning methods. This
study proposes an interpretable and computationally efficient deep learning pipeline combining Principal
Component Analysis (PCA) for feature reduction and a Convolutional Neural Network (CNN) for
classification of ECG signals.
Using data from the publicly available SaMi-Trop cohort, the ECG signals were preprocessed with
a Butterworth high-pass filter to remove baseline wander, followed by PCA to reduce feature space while
preserving critical variance. A custom CNN was then trained on the reduced feature set to classify signals
as Chagas or non-Chagas.
The PCA+CNN model achieved 95.52% sensitivity, 91.11% specificity, 93.31% accuracy, and an
AUC of 96.78%, outperforming the baseline CNN trained on unreduced data. The approach reduced
overfitting, improved generalization, and accelerated convergence, demonstrating strong potential for use
in real-time and low-resource healthcare settings.
