ARRHYTHMIA CLASSIFICATION USING ECG AND PPG SIGNALS WITH CONVOLUTIONAL NEURAL NETWORK METHOD

• Main Author: Monika Saphira, Tasya
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/67781
• Institution: Institut Teknologi Bandung

Arrhythmia is a condition where the heart beats in an irregular rhythm due to abnormalities in its electrical impulse transmission. In this research, convolutional neural network (CNN) is used for arrhythmia classification based on ECG and PPG signals. The classification is done according to arrhythmia classes in the PhysioNet 2015 Challenge dataset, which are Asystole, Bradycardia, Tachycardia, Ventricular Tachycardia. The entire Ventricular Fibrillation class was excluded from classification due to poor data quality. An LSTM layer was added to the CNN as a global feature learning unit. Two kinds of experiment were done: without data selection and with data selection. Data selection was performed to remove data with low PSNR values indicating bad quality. Both experiments were done on ECG and PPG signals, ECG signal only, and PPG signal only. Training and evaluation were done with k-fold cross validation scheme with augmentation done to the train and test data separately to improve data balance. Classification from ECG data only was proven to result in better performance than ECG and PPG data due to the poor quality of PPG signals in this dataset. The experiment without data selection resulted in accuracy of 90,90% and a score of 87,89%, while the experiment with data selection resulted in accuracy of 84,02% and a score of 78,80%. More extensive research and experiment are needed to verify model performance in a larger dataset with better quality.