Evaluation of transfer learning to improve arrhythmia classification for a small ECG database

Abstract

Deep learning algorithms automatically extract features from ECG signals, eliminating the manual feature extraction step. Deep learning approaches require extensive data to be trained, and access to an ECG database with a large variety of cardiac rhythms is limited. Transfer learning is a possible solution to improve the results of cardiac rhythms classification in a small database. This work proposes a open-access robust 1D-CNN model to be trained with a public database containing cardiac rhythms with their annotations. This study explores transfer learning in a small database to improve arrhythmia classification tasks. Overall, the 1D-CNN model trained without TL achieved an average accuracy of 91.73 % and F1-score 67.18 %; meanwhile, the 1D-CNN model with TL achieved an average accuracy of 94.40 % and F1-score of 79.72 %. The F1-score has an overall improvement of 12.54 % over the baseline model for rhythm classification. Moreover, this method significantly improved the F1-score precision and recall, making the model trained with transfer learning more relevant and reliable.