Sleep Fragmentation as a Diagnostic Biomarker of Traumatic Brain Injury

Abstract

Sleep disturbances are among the most prevalent and persistent consequences of traumatic brain injury (TBI), yet they remain underutilized as clinical indicators of injury status. In this perspective, we propose that sleep fragmentation-defined as the frequency of transitions between sleep and wakefulness-represents a functional, scalable, and underrecognized diagnostic biomarker of TBI. Drawing on empirical findings from a mouse model of diffuse TBI, we show that summary measures of sleep fragmentation and duration can reliably distinguish injured from uninjured animals using dimensionality reduction and machine learning techniques. Current biomarkers such as glial fibrillary acidic protein and neurofilament light chain provide valuable insights into structural damage but offer limited information about how injury affects behavior and day-to-day function. Sleep-based metrics, by contrast, reflect neural network integrity and capture ongoing physiological disruption. Critically, these metrics can be collected non-invasively, longitudinally, and in real-world settings using actigraphy, making them a practical complement to blood-based diagnostics that require biological sampling and specialized laboratory infrastructure. Our analysis demonstrates that sleep metrics collected over 48 h post-injury-specifically the number of sleep-wake transitions-carry a strong diagnostic signal. Sleep metrics offer a behaviorally grounded complement aligned with the goals of precision medicine and functional assessment. With further validation, these features may also support monitoring recovery or stratifying injury severity. This perspective highlights sleep fragmentation as a non-invasive diagnostic biomarker for TBI with the potential to enhance individualized monitoring and support early detection efforts in both research and clinical settings.

Keywords: biomarker; concussion; mouse; sleep fragmentation; state transitions.

FIG. 1.

Sleep-based features accurately classify TBI versus sham using random forest models across light and dark periods. Post-injury sleep behavior differentiates TBI from sham animals with high accuracy using a random forest classifier trained on 48-h summary metrics of sleep fragmentation and sleep duration. (A) The model, trained on 80% of the dataset and tested on the remaining 20% (n = 19), used six z-scored features—mean, standard deviation, and range of both transitions and minutes slept per hour—and correctly classified 8 of 10 TBI and 7 of 9 sham animals. (B) A classifier trained and tested using light-period data also correctly classified 8 of 10 TBI and 7 of 9 sham animals. (C) A model using dark-period data correctly classified 8 of 10 TBI and 6 of 9 sham animals. In all panels, the intensity of orange shading reflects the number of animals correctly or incorrectly classified, with darker shades indicating a higher number of predictions in that cell. TBI, traumatic brain injury.

FIG. 2.

Receiver operating characteristic (ROC) analysis confirms strong classification performance of sleep-based features across light and dark periods. ROC curves illustrate the performance of random forest classifiers trained to distinguish TBI from sham animals based on post-injury sleep behavior. (A) The model using 48-h summary features achieved an area under the curve (AUC) of 0.844, confirming strong discriminative ability. (B) A model restricted to sleep data from the light period yielded an AUC of 0.867, reflecting excellent classification performance when mice are normally asleep. (C) The model using data from the dark period also performed well, with an AUC of 0.800, indicating that sleep disturbances during the active phase retain diagnostic value, though with slightly reduced sensitivity–specificity balance. Higher AUC values indicate better model performance across classification thresholds. Results support the utility of both rest- and active-phase sleep metrics for identifying TBI status based on non-invasive measures of sleep fragmentation and sleep loss. A dip in the ROC curve near the midrange threshold likely reflects instability due to the limited test set size and thresholding artifacts common in small-sample classification models. TBI, traumatic brain injury.

FIG. 3.

Sleep fragmentation is the strongest predictor of TBI classification performance. Variable importance analysis from the random forest models revealed that the mean number of sleep–wake transitions per hour was the most influential feature for distinguishing TBI from sham animals. (A) While transitions were the top predictor overall, (B) light-period models, which capture the normal sleep phase, showed that all three transition metrics were the most informative. (C) In contrast, during the dark period, when mice are typically awake, the model also relied on mean minutes slept and standard deviation of minutes slept, suggesting that abnormal sleep quantity during the active phase also contributed to classification. TBI, traumatic brain injury.

FIG. 4.

Principal component analysis (PCA) captures distinct sleep–wake disturbances after TBI across light and dark periods. PCA of post-injury sleep behavior reveals separation between TBI and sham animals. Each analysis used six z-scored summary features over 48 h post-injury. (A) PCA of 48-h sleep behavior shows clear separation between TBI and sham groups along PC1, which captures the largest share of variance and is driven primarily by sleep–wake transition metrics. TBI animals exhibit higher and more variable PC1 scores than sham controls. (B) PCA of light-period data also demonstrated group separation along PC1. PC2 did not differ between groups, supporting the interpretation that sleep fragmentation during the rest phase contributes strongly to group-level differences. (C) PCA of dark-period data similarly revealed significant separation along PC1, while PC2 showed no statistically significant difference. PC, principal component; TBI, traumatic brain injury.