Aims: Sleep-related breathing disorders (SBD) are found in 6-17% of the global adult population and are associated with a high risk of complications such as strokes or type-II diabetes. Previous studies have proposed a real-time finite state machine (FSM) respiratory event detector using as input the nasal pressure signal. This study aims at improving this respiratory event detection by using a light Gated Recurrent Unit (GRU) neural network (48288 parameters).
Methods: This work is based on data from a retrospective 5 centers study (HYPNOS) including 8 hours manually annotated polysomnography of 29 severe obstructive sleep apnea patients. A total of 7846 events were recorded. GRU hyper-parameters were initially optimized using data from 4 patients. A Monte-Carlo approach was applied to generate 100 realizations of the training and testing phases. Each training phase used 20 randomly chosen patients, while the testing phase used the remaining 9. The GRU was trained to classify in real-time each sample of an 8 Hz nasal pressure signal (4.6M samples) as either a normal or respiratory event (apnea or hypopnea). The GRU performances were compared to those reported for the FSM detector.
Results: Apnea and hypopnea detection provided by the GRU detector shows improved performances on the testing database on averaged sensitivity (apnea: 83.1% vs 66.9%, hypopnea: 66.0% vs 47.1%) while preserving a similar averaged positive predictive value (PPV) (apnea: 78.7% vs 80.7%, hypopnea: 35.2% vs 25.0%). The Monte-Carlo approach shows that the GRU performances are characterized by a moderate standard-deviation for apnea (sensitivity: 8.8%, PPV: 6.6%) and hypopnea (sensitivity: 8.1%, PPV: 7.0%) detection on the testing database.
Discussion: The GRU architecture appears as a promising tool for real-time detection of sleep apnea events. Further experiments on a larger database are warranted.