Neurosurgery 99:731–737, 2026
Intermittent neurological examinations can miss gradual or sudden change between bedside assessments, particularly where neurocritical-care expertise is limited. This proof-of-concept study applies computer-vision pose estimation to routine patient video and derives a continuous movement index. Movement increased markedly across higher Glasgow Coma Scale categories and was greater in awake or agitated than sedated patients. The method is noncontact and potentially scalable, but it measures motion rather than neurological function itself. Its clinical value will depend on prospective validation of change detection, false alarms, privacy safeguards and performance across diverse patients.
Objective
To externally compare two artificial-intelligence pose-estimation models, develop a video-derived movement index and test whether that index correlates with bedside measures of consciousness in neuroscience intensive-care patients.
Methods
The investigators retrospectively collected video from patients undergoing video EEG in a large urban neuroscience ICU between July 2024 and January 2025. ViTPose and Meta Sapiens were externally evaluated for pose estimation. The better-performing model was used to construct a computer-vision movement index, which was compared with hourly Glasgow Coma Scale and Richmond Agitation-Sedation Scale assessments.
Main results
The data set contained 998,520 video minutes from 119 patients. ViTPose outperformed Sapiens and was therefore used for the movement index.
Movement rose with increasing GCS: index values were 0.52 for GCS 3–8, 0.70 for GCS 9–13, 3.52 for GCS 14 and 10.99 for GCS 15 (p=0.01), a 21-fold difference between the lowest and highest categories. Patients who were awake or agitated (RASS >−1) had approximately 10 times more movement than asleep or sedated patients (6.59 versus 0.67; p=0.005).
Interpretation
Video pose estimation can convert continuous, otherwise unused observation into a quantitative physiological signal. Correlation with GCS and RASS establishes construct validity, not diagnostic accuracy. Sedation, restraints, weakness, seizures, nursing care and camera visibility can all alter motion independently of global neurological deterioration.
Limitations
This was a retrospective, single-center proof-of-concept study of only 119 patients selected because video EEG was available. It did not test prospective alerts, prediction of neurological events or improvement in patient outcomes. Comparisons with ordinal bedside scores were cross-sectional, and performance by diagnosis, skin tone, body habitus, occlusion or focal deficit was not established.
Clinical takeaway
AI-derived movement should be regarded as a potential adjunct to—not a replacement for—the neurological examination. Before implementation, systems need prospective event-level validation, transparent alarm thresholds, local bias testing, secure video governance and a workflow specifying who reviews and acts on alerts.
You must be logged in to post a comment.