Sleep Apnea Events Classification from a Dual Accelerometry System Using Deep Learning Models

Sleep apnea syndrome (SAS) is a very common chronic disease characterized by the repetition of abnormal and frequent respiratory events, including central and obstructive apneas or hypopneas. Facing this public health problem, the challenge is twofold: improving early detection with minimally intrusive devices, and helping the diagnosis with an automatic scoring of recordings. In this context, we propose to exploit a dual thoracic and abdominal accelerometry system combined with electrocardiography (ECG) to classify sleep apnea events. Several common deep learning architectures were applied and performances were compared against several configurations of signals including nasal airflow from reference polysomnography (PSG), and according to manual expert annotations from PSG recordings. A Gated Recurrent Unit network was found to be the most efficient model and the configuration including ECG and abdominal and thoracic respiratory efforts from accelerometers as input channels provided very promising classification performances of normal and abnormal respiratory events. Thus, the simple technological proposition of dual accelerometry offers the possibility of an automatic identification of SAS events, very close to the expert annotations established from PSG multi-sensors.