0515 Smartphone Multimodal Screening for Obstructive Sleep Apnea

Abstract Introduction Smartphone-based multimodal digital biomarkers may provide an accessible and low-cost alternative to traditional screening of obstructive sleep apnea (OSA) patients. We developed a smartphone app that integrates facial imaging, video-based functional tasks (speech, swallowing, blinking), cardiopulmonary features extracted during paced breathing, and targeted questions. This pilot analysis examined the accuracy of this app to detect OSA risk. Methods The app consists of four components designed based on existing or emerging technologies with documented relevance for OSA detection. Screening questionnaires identify individuals with symptom profiles or risk factors associated with OSA. Facial imaging captures features in facial shape, mandibular structure, and neck morphology that may correlate with upper-airway collapsibility. Video recordings during speech, blinking, and swallowing assess visual and acoustic patterns linked to neuromuscular and physiological traits associated with OSA. Heart rate variability during controlled breathing serves as a surrogate marker of autonomic activity, which is often altered in individuals with OSA. The app was tested in awake participants referred for suspected OSA before undergoing overnight sleep testing across multiple clinical centers. Results More than 585 participants from Australia and the United States have been enrolled, and 497 have completed the assessment and home sleep test (mean age 48 ± 16.3, 38% female, 59% non-Hispanic white, mean AHI 22.3 ± 22.9). Preliminary modelling estimates (n = 209 participants) combining the four multimodal components suggests potential diagnostic performance of 83% sensitivity, 83% specificity, and AUC-ROC 0.85 for detecting OSA defined as AHI ≥ 5. Conclusion This smartphone-based multimodal assessment demonstrates promising potential for identifying OSA using non-contact, easily deployable digital biomarkers. Additional data from non-clinical populations are required to improve dataset balance and enhance model generalizability. A larger prospective validation study across diverse settings is warranted to confirm diagnostic performance and real-world application in clinical settings and at home. Support (if any) Funded by Resmed