Multi-label Arrhythmia Classification From 12-Lead Electrocardiograms

Po-Ya Hsu1, Po-Han Hsu1, Tsung-Han Lee1, Hsin-Li Liu2
1University of California San Diego, 2Central Taiwan University of Science and Techonology


Aims: In this study, we explore the efficacy of signal processing techniques and deep learning (DL) methods for classifying multi-label arrhythmias. Methods: We developed a signal processing and classification strategy to address sample length imbalance and heart rate variance (HRV) among individuals. Such approach consisted of five steps: 1) cleaning raw signals, 2) detecting electrocardiograms (ECG) R-peaks, 3) scaling signals, 4) building DL models, and 5) classifying arrhythmias. First, we remove noise in ECGs by using a low-pass filter with 25Hz cut-off frequency and signal detrending. For R-peaks detection, we ran the Pan-Tompkin’s algorithm on lead II data. If the resulting heart rate is not within 0.5-3.0 Hz, we select R-peaks as the maximum within a sliding window of 1s length and 0.3s stride starting from previous detected R-peak. Next, we address the HRV issue by scaling the signal length with a factor of estimated heart rate divided by sampling frequency. We select residual neural network ResNet-18 as our DL model. The model’s input and output are processed 12-lead ECGs and arrhythmia labels. We experimented with two designs: single model for multi-labels and multiple models with each for one label. We also investigated single versus multiple QRS complexes. We performed cross-validation to tune the optimal scoring threshold, and the training/testing ratios were 80%/20%. Results: We show that single classification model using scaled multiple QRS complexes achieved the best performance. Table 1 displays the performance of five cross-validation tests and submissions. Our best model had F2 scores of 0.983 and 0.726, and G2 scores of 0.726 and 0.476 in cross-validation and submission. We also discovered the discrepancy between cross-validation and submission the lowest in multiple models using scaled multiple QRS complexes and low threshold. Future work: We plan to incorporate cross-channel features, bad channel removal, and wavelet transforms.