Deep LSTM for Personalized ECG Classification Based on Wearable Devices

Jianqiao Zhou, Xin Deng,Kaiwei Sun, Liang Wang, and Boxian Zhang, Key Laboratory of Data Engineering and Visual Computing, School of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Chongqing 400065, China

ABSTRACT

A transfer-learned, deep LSTM model is proposed as a computationally efficient and patient10 specific ECG anomaly detector to be used on wearable devices. A general model was developed using normal sinus ECG data from 18 healthy patients. Then, using transfer learning techniques and a short segment of patient-specific normal sinus ECG data, the general model was retrained to produce a personalized model. The transfer learning model produces superior weighted F2 scores compared to the general model for detecting ECG anomalies such as premature ventricular contractions (n=20, p=.002) and atrial premature beats (n=8, p=.018). A naive model trained only on the retraining data was found to train more slowly and perform worse than the transfer learning model for both premature ventricular contractions (n=20, p=.001) and atrial premature beats (n=8, p=.012). The transfer learning model proposed here provides a sensitive, personalized ECG anomaly detection mechanism that has the added benefit of a lighter, quicker, and more robust learning process to be used on wearable devices.

KEYWORDS

Electrocardiography, Patient monitoring, Personalized medicine, Recurrent neural networks, Wearable sensors.