Development and Validation of a Multimodal Deep Learning Model Using Peripheral Oxygen Saturation and Ballistocardiography Signals for Predicting Apnea-Hypopnea Index in Obstructive Sleep Apnea

This is a prospective observational cohort study designed to develop and validate a multimodal deep learning-based model for estimating obstructive sleep apnea (OSA) severity. Adult participants with suspected OSA who are scheduled to undergo overnight in-laboratory polysomnography (PSG) will be recruited at Fu Jen Catholic University Hospital.

During the PSG examination, participants will simultaneously undergo non-invasive physiological signal acquisition using a smart mattress system equipped with ballistocardiography (BCG) sensors, along with continuous peripheral oxygen saturation (SpO₂) monitoring. These signals will be time-synchronized with PSG recordings. No experimental intervention or alteration to standard clinical care will be performed.

Multimodal deep learning algorithms incorporating convolutional neural networks and sequence-based architectures will be developed to analyze SpO₂ and BCG signals and to estimate the apnea-hypopnea index (AHI). Model performance will be assessed by comparing predicted AHI values with PSG-derived AHI using predefined statistical metrics.

The primary objective is to evaluate the accuracy of AHI prediction using the proposed multimodal model. Secondary objectives include assessment of model discrimination for OSA severity classification and exploratory analyses of physiological signal features associated with sleep-disordered breathing. All analyses will be conducted using de-identified data. This study does not involve U.S. FDA-regulated drugs or devices and is not conducted under an investigational new drug application or investigational device exemption.