O037 Deep learning model of respiratory waveform prediction for an optogenetic therapy for obstructive sleep apnoea.n

Abstract Introduction A novel therapeutic application of optogenetic stimulation as a treatment for Obstructive Sleep Apnea (OSA) requires stimulation of the pharyngeal muscles synchronously with the onset of inspiration to maintain airway patency during sleep1. A sensor array in a removable oral appliance was developed to monitor physiological data; photoplethysmography, acceleration, temperature and audio to predict respiratory airflow. This could then deliver correctly timed stimulation pulses coinciding with inspiration onset as part of a potential therapy. Methods Twenty participants underwent daytime respiratory monitoring (flow, volume and mask pressure) with concurrent monitoring using the oral sensor appliance. Participants were coached to perform multiple breathing patterns; spontaneous, fast and slow breathing, apneas, coughs and snores in supine, left and right-side sleeping positions. Sensor data was used to train a Long Short-Term Memory (LSTM) deep learning model against airflow data to predict respiration waveforms. Results A LSTM deep learning model was trained on over 24,000 seconds of physiological data from the 20 participants to predict respiratory waveforms. The model successfully predicted 5-second respiratory waveforms from the preceding 10 seconds of physiological data with acceptable accuracy (Mean Squared Error = 0.078, Root Mean Squared Error = 0.278), enabling real-time targeted optogenetic stimulation timing. Discussion This study establishes a proof of concept for using an oral sensory appliance in conjunction with optogenetic therapy for OSA. Furthermore, the sensor array demonstrates promise as a potential platform for vital sign and sleep monitoring within the oral cavity.