Research progress and challenges in deep learning for obstructive sleep apnea recognition

Obstructive sleep apnea (OSA) is highly prevalent and harmful; however, its diagnosis relies on polysomnography, which cannot meet the demand for large-scale screening. OSA can cause metabolic abnormalities that in turn drive quantifiable phenotypic changes in facial morphology, voice characteristics, and respiratory patterns, providing a biological basis for noninvasive identification. This review summarizes deep learning-based methods for identifying OSA using oximetry/photoplethysmography, facial images, voice and respiratory sounds, and multimodal data fusion. It analyzes the core challenges currently faced by these models, including insufficient validation, disease coverage bias, lack of interpretability, and technical and ethical constraints on home deployment, and envisions an interpretable, lightweight identification model to promote the clinical translation of home-based OSA screening.