1005 Improved Sleep Apnea Screening After Stroke Using Multimodal Wearable Sensors and Machine Learning

Abstract Introduction Sleep apnea affects over half of stroke survivors, impacting post-stroke recovery and outcomes. Early intervention through continuous positive airway pressure (CPAP) can reduce the impact of sleep apnea, motivating the need for accurate sleep apnea screening tools. Current screening tools such as home sleep apnea tests (HSATs) are difficult to use and uncomfortable due to bulky acquisition systems, electrodes, straps, and/or nasal cannulas, whereas questionnaires (i.e., STOP, STOP-BANG, Berlin OSA) suffer from lack of reliability and low sensitivity. Wireless, multimodal wearable sensors are promising tools to address these gaps; however, stroke-specific algorithms are essential to account for post-stroke physiological changes. The objective of this study was to develop and validate an apnea screening system using wearable sensors and machine learning during early stroke recovery. Methods Patients wore an HSAT (ApneaLink Air) and the Advanced NeoNatal Epidermal (ANNE) Chest and Limb sensors for one night within a week of admission to an acute rehabilitation facility. ANNE sensors are flexible, light-weight (< 17 grams), adhesive based sensors capable of continuous monitoring (up to 7 days). Sleep-based features derived from physiological signals were extracted from ANNE sensors and paired with the ground truth apnea indication from the ApneaLink Air. A logistic regression machine learning model was trained to classify no apnea (normal, AHI< 5) from apnea (mild, moderate, severe, AHI≥5). Results Forty-seven patients were screened with ApneaLink; over 60% had sleep apnea. Models using data from the ANNE Limb sensor performed best, with an F1 score (F1=0.75) outperforming the STOP (F1=0.62), STOP-BANG (F1=0.63), Berlin-OSA (F1=0.62) and a naive classifier (F1=0.38). Model performance increased when distinguishing between AHI< 5 versus AHI≥15 (F1=0.94), whereas the surveys and naive classifier (F1=0.67, F1=0.58, F1=0.58, F1=0.33, respectively) saw a decline in performance. Conclusion Wearable sensors and stroke-specific machine learning models outperformed traditional questionnaires for sleep apnea screening, approaching HSAT-level accuracy (ApneaLink Air). The accessibility, comfort, and ease of use of sensors are well-suited for early application and continuous sleep monitoring compared to commercial HSAT devices. Support (if any) NIH R01HD097786-01A1.