An incremental learning system for atrial fibrillation detection based on transfer learning and active learning.

Background And Objective: Atrial fibrillation (AF) is a type of arrhythmia with high incidence. Automatic AF detection methods have been studied in previous works. However, a model cannot be used all the time without any improvement. And updating model requires adequate data and cost. Therefore, this study aims at finding a low-cost way to choose learning samples and developing an incremental learning system for AF detection.

Methods: Based on transfer learning and active learning, this paper proposed a loop-locked framework integrating AF diagnose, label query, and model fine-tuning. In the pre-training stage, a novel multiple-input deep neural network (MIDNN) is pre-trained using labeled samples from an original training set. In practical application, the model can be used for AF detection. Meanwhile, continuous data is collected to form the candidate set. In the incremental learning stage, the model was fine-tuned continuously by the most informative samples in the candidate set. These samples are selected from the candidate set based on the pre-trained model and a new active learning strategy. The strategy combines the features and the uncertainty of the predicted results.

Results: In order to evaluate the method, the MIT-BIH atrial fibrillation database was used for pre-training and samples of the MIT-BIH arrhythmia database were taken as candidate set. The initial values of Acc, Sen, and PPV were 87.40%, 97.46%, and 81.11%. These indexes reached to the top values of 97.53%, 100.00%, and 95.29% after 14 iterations. Hence, the number of queries was saved by 90.67%.

Conclusions: The proposed system is able to update the model continuously and reduce the labeling cost over 90%. The comparisons demonstrated the effectiveness of MIDNN model and the suitability of novel learning strategy for AF. Moreover, this framework can be extended to other biomedical applications.