Aims: Accurate classification of cardiac rhythms from ECG signals is crucial for early diagnosis of arrhythmias and detection of Chagas disease. Although deep learning models deliver high performance, they often omit essential clinical data and physiological metrics. This study proposes a hybrid approach that integrates clinical and biomedical features with deep representations extracted via a convolutional neural network (CNN).
Methods: ECG signals from the official 2025 Physionet Challenge dataset will be used. Features such as age, gender, heart rate, and the duration and amplitude of P, QRS, and T waves, along with heart rate variability (HRV) metrics, will be extracted using specialized signal processing tools. In parallel, a CNN will process the ECG signals to obtain morphological representations that, when concatenated with the clinical features, are fed into a multilayer perceptron (MLP) for final classification. Frequency domain features and wavelet analysis will also be explored.
Results: Using the basic feature set (age, gender, heart rate, duration, amplitude, and HRV) trained on CODE-15, tests on SaMi and PTB yielded a preliminary challenge score of 0.083, AUROC of 0.656, AUPRC of 0.614, accuracy of 0.573, and F-measure of 0.627. Our best submission on phase 1 obtained a score of 0.376, leaving our team in 56th place out of 178 entries. With additional features, our local results improved 16%.
Conclusion: The results indicate that the basic feature set yields acceptable performance, and we still need to evaluate the use of extra features. Furthermore, the increase in challenge score post-submission may reflect variations in data selection, that still need to be addressed. Future work will integrate CNN-based feature extraction to capture ECG patterns directly and explore detection of specific abnormalities such as AV block, ectopic rhythm or bradycardia, in CODE-15, potentially enhancing early Chagas diagnosis..