CLASSIFICATION OF ARRHYTHMIA UTILIZING MOBILENET-BASED DEEP LEARNING METHOD ON TWO-CHANNELS ECG DATA

• Main Author: Muhammad Nurtsani, Afin
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78864
• Institution: Institut Teknologi Bandung

Sudden Cardiac Death (SCD) is a significant public health issue worldwide, often arising due to severe complications from arrhythmias. One prevalent type of arrhythmia that serves as a reference point for cardiac conditions is Premature Contraction (PC), which is typically categorized into two subtypes: Premature Atrial Contraction (PAC) and Premature Ventricular Contraction (PVC). PACs have the potential to predict mini-strokes, while PVCs can indicate Coronary Heart Disease and tachyarrhythmias.Apart from PC, there's also Bundle Branch Block (BBB), which is divided into Left Bundle Branch Block (LBBB) and Right Bundle Branch Block (RBBB). LBBB is one of the criteria for diagnosing Myocardial Infarction (MI), and RBBB serves as a risk marker for the development of high- degree atrioventricular blocks. The use of a portable Electrocardiogram (ECG), known as a Holter Monitor, is common among doctors for conducting diagnosis related to arrhythmias. To reduce the time cost of observation process, extensive research has been conducted in the development of ECG-based arrhythmia detection systems. Research specifically focusing on detecting Premature Contractions (PC) and Bundle Branch Blocks (BBB) with subject-oriented scheme often employs the MLII channel and provides reasonably good accuracy, although class balance might still be an issue. Certain arrhythmia cases are commonly detected using EKG data from the V1 channel for Right Bundle Branch Block (RBBB), and both the V1 and V6 channels for Left Bundle Branch Block (LBBB). Premature Contractions (PC) patterns are more distinct in the V1 and MLII channels as the P-wave morphology is more pronounced in these channels. The utilization of Convolutional Neural Networks (CNN) with image-based data has emerged as an alternative method for arrhythmia classification. CNNs possess the main advantage of automatically detecting important features and excel in image classification tasks. Therefore, the research aims to perform inter-patient arrhythmia classification using two EKG channels, namely MLII and V1, from the MIT-BIH arrhythmia database. The system developed is specifically designed to detect classes such as normal rhythms, Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), Premature Ventricular Contractions (PVC), and Premature Atrial Contractions (PAC), employing a CNN method based on MobileNet architecture. iv This study aims to develop an arrhythmia classification system using two ECG channels and CNN. Two schemes will be employed: a class-oriented scheme to assess the CNN's classification performance, and a subject-oriented scheme to understand performance variations due to intervariance data. The research process involves collecting EKG signals from the MIT-BIH arrhythmia database. The signals from the database are then subjected to preprocessing, including filtering and segmentation. Filtering is employed to cleanse the data by eliminating noise outside and within the defined cut-off frequencies sequentially. From these processed signals, a classification method is constructed using a parallel architecture of MobileNet. The parallel structure is designed to process image data from the MLII and V1 leads separately, achieved by cropping the top and bottom portions of the signals, which then serve as inputs to their respective MobileNet. The features extracted from the MobileNets are input into an average pooling layer with a size of 2x2. the pooled results are then fed into a fully connected layer and the features from both mlii and v1 leads are combined before entering a dense layer. The final dense layer determines the class based on the extracted features. To comprehensively evaluate the model's performance, all these schemes undergo cross-validation. Based on the conducted research, a MobileNet-based algorithm model has been successfully developed that utilizes image inputs from the MLII and V1 channels for arrhythmia classification. The model performs dominantly in the class-oriented classification scheme and competes well with previous features for BBB and PC arrhythmias. The average accuracy, sensitivity, and specificity values across all classes for this model are 97.9%, 95.1%, and 98.8%, respectively.However, in the subject-oriented scheme, the model's performance in classification is not yet optimal. While the sensitivities for the "normal" and "LBBB" classes exhibit values of 92% and 90% respectively, some classes have lower sensitivity rates compared to other studies. For instance, the "RBBB," "PAC," and "PVC" classes have sensitivities of 47%, 7%, and 70%, respectively. Nevertheless, it's noteworthy that the "normal" class demonstrates a high specificity rate of 79%, which is favorable. Predicting individuals with arrhythmias as normal is more desirable than the opposite scenario.