An embedded system for the classification of sleep disorders using ECG signals

Abstract

Sleep apnea (SA) is a well-known sleep disorder. It predominantly appears due to lack of oxygen in humans. Identifying SA at an early stage can help early diagnosis. The primary motto of our research is to identify SA using electrocardiogram (ECG) signals. Here, three classes are considered for classification. One is normal (N), and the other two are SA classes obstructive sleep apnea (OA) and central sleep apnea (CA). ECG signals are accumulated for MIT-BIH polysomnographic dataset. The ECG data divided into ECG segments and labelled using annotation file. The proposed deep long short-term memory (LSTM) model is then trained using ECG segments and further tested. The model is then finetuned and optimized to obtain the best accuracy. An accuracy of 98.51% is obtained. In addition, performance measures like precision, sensitivity, specificity, F-score are also evaluated. The model is then deployed on NVIDIA’s Jetson nano board to build a prototype. Our model is effective, promising and outperformed existing state of art techniques.