Tendencies of Unsupervised Learning on ECG Signals

This study explores the internal mechanisms of a well-trained convolutional auto encoder applied to electrocar-diogram (ECG) signals, aiming to identify specialized neurons for different wave types, such as T-wave, P-wave and QRS complex. Utilizing ECG data from the LUDB dataset, collected from both healthy individuals and patients with various car-diovascular conditions, we trained and analyzed a convolutional auto encoder. Our findings indicate that while the autoencoder is proficient at distinguishing T-waves from other types of wave, it lacks neurons specialized for QRS complexes and P-waves. This observation was confirmed through Mann-Whitney U tests. It was also verified that when changing the hyperparameters of the convolutional autoencoder, the obtained tendencies are preserved, indicating their stability. The research underscores the potential of unsupervised learning methods in medical signal processing, highlighting both their capabilities and limitations, which contributes to the ongoing effort to develop more effective diagnostic tools for cardiovascular diseases through advanced machine learning techniques.