CNS*2026 Halifax

Introduction
Sleep disturbance is a common consequence of neurotrauma, including traumatic brain injury (TBI) and spinal cord injury (SCI), yet objective biomarkers capable of distinguishing injury type remain limited. Wearable actigraphy generates high-dimensional physiological time-series data that may contain latent signatures of injury-related sleep disruption. Although sleep disturbances also occur following severe orthopedic injury (SOI), it remains unclear whether computational analysis of actigraphy-derived sleep architecture can detect injury-specific phenotypes. Here, we tested whether supervised machine learning models could classify injury type from wrist actigraphy-derived sleep signatures.


Methods
Continuous wrist actigraphy was collected from 61 inpatients (TBI n = 33, SCI n = 12, SOI n = 15). Raw actigraphy time-series were processed using a custom Python-based computational pipeline that performed automated preprocessing, circadian segmentation and high-throughput extraction of sleep–wake metrics. This workflow produced 24 features per observation, including sleep efficiency, fragmentation, and bout structure. Five engineered composite features capturing circadian disruption and sleep consolidation yielded a 31-feature representation of sleep dynamics. Seven supervised machine learning classifiers were evaluated using stratified five-fold cross-validation with performance assessed by AUC and average precision. (Fig. 1)


Results
A stacking ensemble method achieved the strongest performance across both classification tasks. Discrimination was strong for TBI versus SOI (AUC=0.825±0.033; average precision=0.896) and SCI versus TBI (AUC=0.900±0.054; average precision=0.975). Nighttime sleep efficiency emerged as the most informative feature, suggesting that nocturnal sleep consolidation carries strong diagnostic value across injury groups.


Discussion
These findings demonstrate injury-specific sleep signatures and show that actigraphy-derived sleep phenotyping can distinguish central nervous system injury from peripheral trauma using scalable, noninvasive monitoring. This computational framework may enable objective clinical stratification in neurotrauma and provides a foundation for future machine learning approaches to track recovery trajectories and identify sleep-based biomarkers of injury burden.

Figure 1. Actigraphy-based sleep feature pipeline and machine learning model performance for TBI classification.​

References
1. Courville, E., Kazim, S. F., Vellek, J., Tarawneh, O., Stack, J., Roster, K., Roy, J., Schmidt, M., & Bowers, C. (2023). Machine learning algorithms for predicting outcomes of traumatic brain injury: A systematic review and meta-analysis. Surgical Neurology International, 14, 262. https://doi.org/10.25259/SNI_312_2023
2. Tunthanathip, T., & Oearsakul, T. (2021). Application of machine learning to predict the outcome of pediatric traumatic brain injury. Chinese Journal of Traumatology, 24(6), 350–355. https://doi.org/10.1016/j.cjtee.2021.06.003
 

Acknowledgement
Acknowledgements: A.L. was supported by the National Institutes of Health Training Program in Sleep and Circadian Biology (T32HL149646). Additional thanks to the Sleep, Inflammation, and Neuropathology Lab members for the support.