Classification of ECG signals based on deep learning and chromaspectrograms

ECG is a non-invasive diagnostic tool that graphically records the rhythm and electrical activity of the heart, allowing for the evaluation of changes in the heart's electrical activity over time. It contains essential physiological information for analyzing heart function, and is therefore widely used by cardiologists to detect health disorders. However, the ECG signal is non-linear and may contain small changes that are difficult to detect with the naked eye due to its low amplitude. Although arrhythmia does not always directly threaten human health, it can lead to symptoms of heart disease and more severe arrhythmias, such as rapid atrial fibrillation, paroxysmal supraventricular tachycardia, and persistent ventricular tachycardia. Arrhythmias can also cause palpitations, chest tightness, dizziness, changes in blood pressure, sweating, syncope, and sudden death. These heart conditions are dangerous and require urgent treatment if symptomatic. With the development of artificial intelligence, such diseases can be more easily detected. In this study, ECG signals were transformed into chroma spectrograms using deep learning methods to visually represent the harmonic structure of the signals, thereby improving the diagnostic accuracy of the model. Features were then extracted from these signals using pre-trained networks such as InceptionNet V3, GoogleNet, MobileNet V2, EfficientNet, AlexNet, and ResNet-50. All the networks were evaluated using 10-fold cross-validation, as well as 70%-30% and 80%-20% training-testing splits for validation. For the classification process, arrhythmia classification was performed using L-SVM, Q-SVM, C-SVM, KNN, BT, and LR classifiers. The InceptionNet V3 model achieved 84.7% accuracy, the GoogleNet model 85.9%, the MobileNet V2 model 87.6%, the EfficientNet model 87.5%, the AlexNet model 87.3%, and the ResNet-50 model 87.2%.