Atrial fibrillation identification based on a deep transfer learning approach

Abstract

Objective: Detection of Atrial fibrillation (AF); as a very common cardiac arrhythmia; is a challenging issue because it is often asymptotic. Most of previous studies were based on feature extraction or training CNNs from scratch. Difficulties in finding appropriate features, requiring an expertise for feature extraction and tuning parameters, and needs for a large amount of labeled data are the most drawbacks of previous studies. The transfer learning is a solution for these problems especially in the medical analysis where available data is limited. The main goal of this study is to investigate the ability of transfer learning method for classification ECG signals into normal, AF, other rhythms and noisy classes. Approach: In our analysis, ECG signals have been transformed to 6 s segments images and fed to three well-known pre-trained models (AlexNet, VGG-16, and ResNet-152). Features have been extracted from different layers of the models and used as the inputs of a classifier for AF detection. Then we have compared the best resulting model to proposed models trained from scratch. The proposed method also investigated the impact of models depth. The algorithms have been trained and validated using the public dataset of 2017 Physionet challenge. Main results: The method achieved average accuracy 87.9, and F-score 86.4 on the validation dataset (publicly available dataset). The second level of extracted features has the highest accuracy for all the pre-trained models in our study. Pre-trained AlexNet outperformed full training CNNs evaluated in this study. Significance: It is shown that the transfer learning method yields good accuracy recognition performance even when source and target datasets are completely different. The results show that the transfer learning is a reliable, accurate and low computational method to classify AF from short ECG signals while requires neither feature selection nor heavy processing steps.