Context: Chagas disease affects around 8 million people worldwide, being one of the main causes of cardiomyopathies. In patients with structural heart disease, echocardiography (ECO) provides important implications for diagnosis, prognosis, and treatment. However, ECO is more expensive, complex and difficult to access than electrocardiography (ECG).
Objective: Verify the feasibility of predicting a reduced left ventricular ejection fraction (LVEF), a metric estimated by ECO, using ECG signals from patients with Chagas heart disease (ChHD).
Data description: We used ECG holter signals from 219 ChHD patients obtained from University Hospital Clementino Fraga Filho – Federal University of Rio de Janeiro, Rio de Janeiro, Brazil. The local ethics committee approved the research (number 45360915.1.1001.5262).
Data preprocessing: Initially, we combined Wavelet and Hilbert transforms and mathematical models for detecting and delineating each QRS complex, P and T waves. Then, we defined a beat window (BW) as starting at each P-wave onset and finishing at each subsequent T-wave end. Next, we removed BWs whose length exceeded the lower or upper bounds of Turkey boxplot, and then, we applied zero-padding to make the remaining BWs the same length.
Data Processing: A Long Short-term Memory (LSTM) network was trained with sets of 200 BWs, aiming to classify normal BWs (LVEF ≥ 0.5), and anomalous BWs (LVEF < 0.5). The hyperparameters of the LSTM network were chosen by the Bayesian Optimization algorithm. The final structure of the network consists of 3 LSTM layers (each one containing 100 units), followed by 4 dense layers (with 64, 32, 16 and 8 neurons, respectively).
Results: The LSTM model got an accuracy of 0.66 and a F1-score of 0.66.
Conclusion: The preliminary results evidence the potential of developing models to predict heart failure from single-channel ECG signals obtained from ChHD patients with relatively short time duration.