Prediction of Left Ventricular Ejection Fraction Using an ECG-based LSTM Model in Chagas Disease Patients

Objective: Classify the left ventricular ejection fraction of chagasic patients into preserved and non-preserved by using electrocardiogram signals. Context: Left ventricular ejection fraction is an important indicator of heart failure and predictor of sudden death. To estimate this indicator, echocardiography is necessary, which is usually more expensive and restrictive than electrocardiography. Methods: Initially, we separated the signals into two classes: ejection fraction less than 0.5 (class 1) and ejection fraction greater than or equal to 0.5 (class 2). We used a Tukey's boxplot to separate noisy beats from non-noisy ones based on their duration. Next, we applied an LSTM (Long-Short Term Memory) network to classify sets of 200 beats of each signal. Finally, we applied an artificial neural network to obtain a class for the entire signal, using the LSTM outputs of each set of 200 beats. Results: We obtained, as the best result, an accuracy of 0.79 and a F_{1} - score of 0.78. Conclusion: We obtained satisfactory results. However, we believe that they can be improved by a more sophisticated beat selection method and a more robust LSTM model.