Early Recognition of Sudden Cardiac Death Combined with Deep Transfer Learning and Superficial ECG Features

Sudden cardiac death (SCD) is a global public safety issue that poses a serious threat to human life and health. Moreover, the clinical records of SCD data are extremely limited. This paper proposed an early prediction algorithm for SCD based on deep transfer learning. The algorithm used a two-stage training approach, which included pre-training and fine-tuning of convolutional neural networks (CNNs) to automatically capture the deep heart rate features (HRFs) in high-risk electrocardiogram (ECG) signals before SCDs. The aim is to explore automatic early prediction of SCD under limited ECG data training. This paper extracted three HRFs as input sets from 30 second single lead ECG signals of 18 patients with malignant arrhythmia and 18 healthy subjects, including RR interval, dRR interval, and turbulence onset (TO). Then, the ten-fold crossover method was used to evaluate the classification performance of this model. Finally, the results showed that the average accuracy (Acc), sensitivity (Se), and specificity (Sp) of predicting SCD within 30 minutes before onset were 91.79%, 87.00%, and 96.63%, respectively. The transfer learning algorithm proposed in this paper effectively addresses the issue of insufficient data in the early prediction models of SCD. It also achieves early and accurate prediction of high-risk signs of SCD by using ECG. This is expected to form an intelligent and effective monitoring and early warning technology for high-risk patients of SCD.