TWO-STAGE MODIFIED ENCODER FOR ECG HEARTBEAT CLASSIFICATION VIA FUSING MORPHOLOGICAL AND TEMPORAL FEATURES

• Main Author: hermawan, Agus
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/71775
• Institution: Institut Teknologi Bandung

Cardiovascular Diseases (CVD) are the leading cause of death in Indonesia and World. In Indonesia, the number of expert (cardiologists) limited and scattered unevenly. In the medical field, an electrocardiogram (ECG) is the gold standard for detecting heart disease. The focus of this study was to build an Artificial Neural Network-based model capable of accurately classifying arrhythmias from ECG signals. Physionet MIT-BIH (The Massachusetts Institute of Technology – Beth Israel Hospital) Database used to train and evaluate the model. A Two-Stage Modified Encoder (TSME) is proposed as a model for the arrhythmia classification system. This model uses two inputs: 187 samples were extracted from morphological waveforms, and 8 temporal features were calculated from the R-R interval of heartbeats. Input from morphological waveforms processed with convolution block and attention block, then concatenated with input from temporal features. The proposed method was evaluated with inter-patient paradigm to conform to the AAMI standard, where the subject (record) differentiated during the training and testing phases. The obtained sensitivities of Normal (N), Supraventricular (S), Ventricular (V) and Fusion (F) were 94.24%, 90.02%, 92.89% and 18.56%. Positive predictivity values of the model are 99.79%, 43.55%, 88.59% and 14.49% for N, S, V and F class. The overall accuracy of four classes of heartbeats is 93.41%. The comparison shows that our proposed methods have achieved competitive performance results compared to state-of-the-art. To assess how well the model generalized the problem, the INCART Database was evaluated as a different source of data. The method was suitable for clinical application as both a high positive predictive value for the N class and high sensitivity for the abnormal classes (S and V).