Machine Learning Classification of Traumatic Brain Injury Using Video-Based Pupillometry

-Introduction: Traumatic brain injury (TBI) frequently disrupts autonomic function, which may manifest as altered pupillary dynamics. While manual pupillary assessment remains standard in acute care, quantitative video-based pupillometry is underutilized for TBI characterization, and objective automated analysis tools are lacking.

Objective: The objective of this study is to develop and validate a fully automated pipeline that classifies TBI versus healthy controls using passive, task-free video-based pupillometry and to identify which specific dynamics provide the most discriminative information.

Methods: We recorded infrared video from 55 participants (34 TBI and 20 healthy controls; 98 recordings total) under resting conditions. A complete automated preprocessing pipeline (ROI detection, artifact removal via median absolute deviation, and linear interpolation) extracted a comprehensive set of time-series features, including statistical moments, velocity, and fluctuation. After feature standardization and univariate selection (ANOVA F-value), an SVM with an RBF kernel was optimized via grid search and evaluated using stratified 5-, 7-, and 10-fold cross-validation.

Results: The optimized SVM achieved a mean cross-validated accuracy of 86.73% (AUC = 0.944, Cohen's κ = 0.709, sensitivity: 89.06%, and specificity: 82.35%). Feature ranking revealed that variances in pupil area (reflecting spontaneous hippus), maximum constriction velocity, area distribution skewness, recovery slope, and inter-pupillary asymmetry were the most informative metrics—pointing to autonomic lability, midbrain pathway integrity, and bilateral symmetry as key discriminating dimensions. Performance remained stable across all cross-validation schemes, with learning-curve analysis indicating a modest bias–variance gap (8.2%) consistent with good generalization.

Conclusions: This proof-of-concept study demonstrates that passive pupillometry, combined with a validated SVM pipeline, can reliably distinguish TBI from healthy controls using features derived from spontaneous pupil behavior. The fully automated workflow and lack of active patient cooperation make this approach particularly suited for serial monitoring and rapid triage. Although external validation on larger, multi-center cohorts is needed, our findings establish a foundation for objective, scalable screening in autonomic neurology.