Pose AI prediction of neurological status in the Neuroscience Intensive Care Unit
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
BACKGROUND
The neurological exam is pivotal in assessing patients with neurological conditions but has severe limitations: it can vary between examiners, it may not discern subtle or subacute changes, and because it is intermittent can delay recognition of new deficits. Even in the Neuroscience Intensive Care Unit (NSICU), neurological exams are conducted only hourly. The majority of ICUs/hospitals lack subspecialized neurocritical care services, exacerbating this neurologic monitoring gap. We hypothesized that pose AI, a machine learning approach to track patient position, could provide a continuous and relevant method of neurological monitoring.
METHODS
We retrospectively collected video segments from patients in the NSICU and the Epilepsy Monitoring Unit (EMU) who underwent video-EEG at a large, urban hospital between July, 2024 to January, 2025. We externally validated two leading pose AI models, ViTPose and Meta Sapiens. We then developed a robust movement index and evaluated its correlation with two measures of consciousness obtained through hourly physical exams, the Glasgow Coma Scale (GCS) and Richmond Agitation Sedation Scale (RASS).
RESULTS
We collected 998,520 video minutes from 119 NSICU and EMU patients. ViTPose demonstrated superior performance to Sapiens across multiple metrics, so we used ViTPose to calculate a computer vision movement index (λ MI ). We observed higher movement with increasing GCS (GCS 3–8 λ MI =0.52, GCS 9–13 λ MI =0.70, GCS 14 λ MI =3.52, GCS 15 λ MI =10.99, P =0.01), a 21-fold increase from the lowest to highest tranche. We also observed 10-fold higher movement in awake/agitated patients (RASS>-1 λ MI =6.59) compared to those who were asleep/sedated (RASS≤-1 λ MI =0.67, P =0.005). Taken together, we developed a novel computer vision movement index and demonstrated expected correlations with GCS and RASS scores in NSICU patients.
CONCLUSION
We show that pose AI can provide minimally invasive, continuous and clinically relevant neuro-monitoring in critically ill patients. Neurological conditions account for the highest global disease burden and pose AI may be a low-cost, explainable, and scalable AI solution to address this pressing need for neuro-telemetry.
