Multimodal Deep Learning Model for Predicting the Apnea-Hypopnea Index in Obstructive Sleep

Obstructive sleep apnea is a common sleep disorder closely associated with cardiovascular, metabolic, and neuropsychiatric comorbidities. It is characterized by repeated upper airway collapse during sleep, leading to intermittent hypoxia and sleep fragmentation. Although polysomnography remains the diagnostic gold standard for obstructive sleep apnea, its high cost, complexity, and limited accessibility pose challenges for large-scale screening and early identification. Recent advancements in noninvasive sensing technologies-such as electronic stethoscopes, wearable oximeters, and under-mattress pressure sensors-have enabled low-burden physiological monitoring solutions, offering new opportunities for simplified obstructive sleep apnea detection. In this study, synchronized multimodal physiological data will be collected during overnight sleep, including respiratory sounds, continuous saturation measurements, and standard polysomnography waveforms. Signal preprocessing and feature extraction will be performed to ensure data quality and temporal alignment. A deep learning model will be developed using these multimodal signals as inputs. The apnea-hypopnea index will be derived from overnight polysomnography. The model will be trained to estimate apnea-hypopnea index values and classify obstructive sleep apnea severity according to established clinical thresholds.

Inclusion criteria

• age 30-75 years
• clinically suspected obstructive sleep apnea and scheduled for polysomnography
• willing and able to provide written informed consent

Exclusion criteria

• intolerance to the electronic stethoscope or fingertip pulse oximeter
• significant structural airway abnormalities
• arrhythmia
• neuromuscular disorders
• pregnancy
• hospitalization within the past 1 month
• inability to provide informed consent or requiring legal guardian consent