Transfer Learning for Detection of Atrial Fibrillation in Deterministic Compressive Sensed ECG

Atrial Fibrillation (AF) is a cardiac condition resulting from uncoordinated contraction of the atria which may lead to an increase in the risk of heart attacks, strokes, and death. AF symptoms may go undetected and may require long-term monitoring of electrocardiogram (ECG) to be detected. Long-term ECG monitoring can generate a large amount of data which can increase power, storage, and the wireless transmission bandwidth of monitoring devices. Compressive Sensing (CS) is compression technique at the sampling stage which may save power, storage, and wireless bandwidth of monitoring devices. The reconstruction of compressive sensed ECG is a computationally expensive operation; therefore, detection of AF in compressive sensed ECG is warranted. This paper presents preliminary results of using deep learning to detect AF in deterministic compressive sensed ECG. MobileNetV2 convolutional neural network (CNN) was used in this paper. Transfer learning was utilized to leverage a pre-trained CNN with the final two layers retrained using 24 records from the Long-Term Atrial Fibrillation Database. The Short-Term Fourier Transform was used to generate spectrograms that were fed to the CNN. The CNN was tested on the MIT-BIH Atrial Fibrillation Database at the uncompressed, 50%, 75%, and 95% compressed ECG. The performance of the CNN was evaluated using weighted average precision (AP) and area under the curve (AUC) of the receiver operator curve (ROC). The CNN had AP of 0.80, 0.70, 0.70, and 0.57 at uncompressed, 50%, 75%, and 95% compression levels. The AUC was 0.87, 0.78, 0.79, and 0.75 at each compression level. The preliminary results show promise for using deep learning to detect AF in compressive sensed ECG.