322 Pose Tracking of Patient Video Data in the Neuroscience Intensive Care Unit: A Preliminary Analysis

INTRODUCTION: Patients are routinely sedated in the Neuroscience Intensive Care Unit (NSICU) for procedures, imaging, and clinical pathologies including cerebral edema and elevated intracranial pressure. Titrating sedation is challenging: patients need to be sedated enough to tolerate uncomfortable therapies but not overly sedated to be mistaken for declining neurological status. Thus, accurate and continuous characterization of level of arousal are acornerstone of neurocritical care. Concurrently, there is a vast amount of patient video monitoring data being collected, but it is not routinely analyzed. METHODS: We obtained video-EEG from a patient in the NSICU. We calculated pose with a deep learning computer vision approach, MMPose. As a summary metric of movement, we calculated the standard deviation (SD) in pose within each 10 second interval. We then correlated movement with the Richmond Agitation-Sedation Scale (RASS), a validated sedation scale. Lastly, we trained an XGBoost classifier to predict level of sedation. RESULTS: We compared movement between light and deep sedation (RASS=0 vs RASS=-4). Movement was approximately twice as high with RASS 0 (SD=1.7 pixels) compared to RASS -4 (SD=0.84 pixels, Wilcoxon rank sum test P=1.4x10^-45). Furthermore, movement was higher in the distal extremities compared to the head at RASS 0 (P=1.5x10-^15) but not RASS -4 (P=0.77), further supporting the algorithm’s performance in detecting light versus deep sedation. Sedation prediction with XGBoost had high performance on training and test sets (ROC-AUCs 0.999 and 0.995, respectively). CONCLUSIONS: We used pose tracking to differentiate light versus deep sedation in a patient in the NSICU. This technology may improve sedation titration and, more generally, offer a scalable, minimally invasive tool for monitoring all patients with acute neurological illness.