ARRHYTHMIA CLASSIFICATION BASED ON TWO LEAD ECG USING MACHINE LEARNING

• Main Author: Sari Hayunah Nurdiniyah, Elsa
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/67552
• Institution: Institut Teknologi Bandung

Arrhythmia is a heart disorder in which the heart beats abnormally. Arrhythmias can cause symptoms such as fatigue and chest pain. The most severe conditions can lead to stroke and heart failure, leading to death. ECG is usually used as a tool for the detection of arrhythmias. In arrhythmia research, the inter-patient method is needed to obtain unbiased results. However, most studies using the intra-patient method have detected large classes of arrhythmias such as normal, ventricular arrhythmias, atrial arrhythmias, and fusion beats. Not many inter-patient studies have detected more specific arrhythmias such as bundle branch block and premature beats. Several studies that detect a specific type of arrhythmia with interpatients use only one ECG channel (MLII) as research data. The classification results show a fairly good accuracy even though it is not balanced for each class. The MLII channel is widely used as research data for arrhythmias because the electrical activity recorded by the MLII is the same as the direction of conduction of heart impulses. However, some arrhythmias such as LBBB and RBBB are generally detected using the ECG of the V1 and V6 channels. In addition, the ectopic beat (PAC, PVC) is more visible in the V1 and MLII channels because the P waveform is more visible in both channels. In the classification of arrhythmias, machine learning is used to help classify arrhythmias more quickly. Some machine learning techniques include SVM, random forest, KNN, etc. In addition, there is also ensemble learning which is a combination of several machine learning methods. Ensemble learning has the advantage of producing a general and not overfit classification method, which is very much needed in classifying arrhythmias. Therefore, this study will classify inter-patient arrhythmias using two ECG channels, namely the upper and lower channels from the MIT-BIH arrhythmia database. The system made specifically detects normal classes, left bundle branch block (LBBB), right bundle branch block (RBBB), premature ventricular contractions (PVC), and premature ventricular contractions (PAC) using machine learning methods. This study aims to build a classification system for arrhythmias using two ECG channels and machine learning with the inter-patient method. Besides, this study aims to evaluate the effect of using two ECG channels on the results of classification performance compared to other studies. The research process includes taking ECG signals from the MIT-BIH arrhythmia database. The signal from the database is then cleaned using a median filter and a low pass filter. The cleaned signal is then segmented around the R peak from the beginning of the P wave to the end of the T iv wave. Then the segmented ECG signal is extracted using the TSFEL function package. The extracted features include statistical features, temporal features, and spectral features. In addition to the three, the interval R feature is also used. Before the feature is selected, the data is separated into training and test data, so biased results are not obtained. The feature selection process will be based on the training data's features. The features used in the test data will follow the features selected from the training data after the feature selection process. The feature selection process is based on the label's correlation value and the correlation with other features. Features that have a high correlation with the label and have a low correlation with other features will be used. Furthermore, the training and validation are carried out using training data to get the best parameters. From these results, a classification method is built using a combination of various machine learning (ensemble learning) parameters selected from the tuning results. The best outcome will be selected as the method of arrhythmia classification. Based on the research conducted, a machine learning model was successfully built from a combination of K-nearest Neighbors, Random Forest, and Support Vector Machine with two ECG channels, namely the MLII-V1 channel. The weight combination that produces the best performance for each machine learning is 2,2,1. The results of this combination show good results compared to other similar studies, where in this study, an accuracy of 87%, recall 87%, precision 88%, and an f1 score of 87% was obtained. The use of multi-channel, namely the MLII-V1 channel, was proven to improve the performance of the five-class arrhythmia classification. This is indicated by the comparison of higher classification results with single-channel configurations and with the results of other studies that only use single-channel.