A photoplethysmography-based diagnostic support system for obstructive sleep apnea using deep learning approaches

Abstract

Obstructive Sleep Apnea (OSA) is a common sleep disorder characterized by periods of reduced or complete cessation of airflow during sleep due to obstruction of the upper respiratory pathway. A novel deep learning framework is developed for automated feature extraction and detection of OSA events from Photoplethysmogram (PPG) signals recorded at the finger tip of the subjects using a Photoplethysmography sensor. This helps in real-time automatic OSA screening at a faster rate and reduces the need for an exhausting and time-consuming Polysomnography (PSG) sleep study. Bi-directional Long Short-Term Memory (Bi-LSTM), Temporal Convolutional Network (TCN), and TCN-LSTM are the three deep learning approaches implemented to facilitate the automatic screening of OSA events, and their performance is compared. Training and testing are carried out using datasets collected from Physionet's apnea database and real time PPG signals of 315 subjects from diverse age groups with health conditions viz., hypertension, cardiovascular disease, and OSA. The performance of TCN-LSTM is better compared to the performance of TCN and Bi-LSTM. The proposed system exhibits an accuracy of 93.39%, a specificity of 94.37%, a sensitivity of 98.98% and F1 Score of 94.12%.