Classification of Arrhythmias Using a Pre-trained Deep Learning Model with Binary Images of Segmented ECG

Introduction . Arrhythmia or irregular heartbeat occur when the heart’s electrical system is disorganized or out of sync, which may cause strokes, sudden cardiac death, and other complications. The introduction of an automated classification of arrhythmias based on deep learning could facilitate the decision-making process by saving time and labor resources. Aim . To study the performance of a modified arrhythmia classification improved by using binary images of segmented ECG signals with combinations of orthogonal and surface signals. Materials and methods . This article studies an arrhythmia classification based on binary images of surface and orthogonal ECG signals. The data labeling was automated using the Python programming language. Initially, all signals are subjected to preprocessing followed by their plotting and segmenting in 2-second windows. Next, those segments are saved as RGB images followed by their conversion into binary images, where the signal is white, and the background is black. Finally, the pre-trained Alexnet model is used to classify nine classes, where each surface ECG and orthogonal lead is classified separately. Results . The performance of the model is evaluated by the mean accuracy, precision, F1-score, and confusion matrix of all leads. The results of a parallel classification of 12 lead ECG are better than those for the orthogonal leads. All leads with accuracy, precision, and F1-score equal to 0.84, 0.78, and 0.71, respectively. Conclusion . The performance of the model was evaluated for three cases: 12 surface ECG leads, orthogonal leads, and all leads. The calculated mean values of accuracy, precision, and F1-score for each case confirmed the sufficiency of the 12-lead surface ECG for classifying nine different types of arrhythmia using binary images of ECG segments.