Multiclass Classification of ECG Signals Using a LSTM Network Trained with Spectral Features

In this work, an implementation of a deep learning algorithm based on a LSTM network trained with spectral estimates obtained from ECG signals is presented, among which are: AR Burg spectrum, fast fourier transform spectrum, welch spectrum, Yule-Walker spectrum, periodogram spectrum, covariance spectrum. The objective of this work is to demonstrate that spectral estimates can generate good results when training an LSTM, taking into account the processing time, computational cost, the complexity of feature extraction and the value of quality metrics and based on the results, this model can be adapted to other types of approaches. The data used for training corresponds to the MITBIH database, which includes normal rhythm, auricular premature contraction, ventricular premature contraction, left bundle branch block and right bundle branch block. Using the signal analysis techniques mentioned above, the data was transformed into sequential data to improve the results, the corresponding features were extracted and a dimensionality reduction was applied. This features were used to train the BiLSTM network and for each train with different features, the results were validated and compared with the quality metrics. The spectral estimate that obtained the best results was FFT with values: 94.6% Accuracy, 94.6% Recall, 94.9% Precision, 94.7% F1-score, 93.4% Matthews and 83.3% Kappa.