Predicting Traumatic Brain Injury Post-Trauma Using Temporal Attention on Sleep-Wake Data

Abstract

Background:

Traumatic Brain Injury (TBI) is a major public health concern, and accurate classification is essential for effective treatment and improved patient outcomes. Sleep/wake behavior has emerged as a potential biomarker for TBI classification, yet the optimal time window in which to identify sleep/wake changes after TBI remains unclear.

Methods:

We evaluated daily longitudinal sleep/wake data from a prospective cohort of more than 2,000 emergency department patients with and without blood biomarker-documented TBI (Glial Fibrillary Acidic Protein - GFAP > $268\frac{pg}{ml}$). We utilized a deep learning model to identify the impact of time from trauma and duration of data collection on the model’s ability to distinguish between TBI-positive (TBI+) and TBI-negative (TBI−) cases.

Results:

Our analysis showed that sleep/wake data from the first 7 days after TBI most accurately identified TBI. Sleep-wake data from the first 7, 14, and 21 days after trauma achieved sensitivity/specificity of 81%/25%, 40%/66%, and 45%/58%, respectively. F1 scores of deep learning models developed from the first 7, 14, and 21 days were 22%, 21%, and 20%, respectively.

Conclusions:

The results suggest that early sleep/wake data has promise for assisting with TBI identification.

Significance:

In the future, the incorporation of sleep/wake derived biomarkers into TBI identification tools could assist in the identification of individuals with potential TBI for further screening and intervention.

I. INTRODUCTION

Traumatic brain injury is a significant global public health concern resulting from external forces that disrupt brain function [1]. The primary causes of TBI include falls, which are the leading cause, especially among older adults and young children. Vehicle-related collisions, such as car, motorcycle, and bicycle accidents, are also common sources of TBIs, particularly among young adults. Additionally, violence, including domestic violence, gunshot wounds, and child abuse, can result in TBI. Sports injuries from contact sports like football, hockey, and boxing pose a high risk for TBIs. Lastly, explosive blasts, particularly relevant for military personnel, can cause TBIs through the pressure waves generated by explosions. In the United States, TBIs are a leading cause of emergency department visits, and also a significant cause of death, accounting for roughly 30% of all injury-related fatalities related to injuries [2], [3]. This highlights the urgent need for effective prevention and treatment strategies. [1].

TBI can vary widely in severity and is typically classified into three categories: mild, moderate, and severe. Each category presents a distinct range of symptoms that presumably reflect the extent of brain damage. Common TBI symptoms encompass cognitive impairments, physical issues, sensory disruptions, and communication difficulties. Among these, sleep-related symptoms are objectively measurable and can significantly impact an individual’s quality of life [4]–[6]. Sleep disturbances in TBI patients often include insomnia [7], [8], characterized by difficulty falling or staying asleep [9]; excessive daytime sleepiness, which leads to challenges in sustaining wakefulness during waking hours; and nightmares that disrupt restful sleep. Additionally, TBI may be complicated by sleep apnea, a condition marked by interrupted breathing during sleep. Understanding and addressing these sleep-related issues is crucial for improving overall recovery and well-being in individuals affected by TBIs.

In recent years, wearable devices have become essential tools for managing a range of health issues. For example, they can monitor heart conditions like arrhythmias and blood pressure fluctuations, track sleep patterns, detect sleep apnea, and much more [10]–[13]. These devices offer low-cost, easy-to-use sensor technologies that enable continuous health monitoring and provide valuable insights into a patient’s well-being. As previously noted, one of the hallmark symptoms of TBI is the presence of sleep disorders, such as difficulty falling or staying asleep. Actigraphy, a method for measuring overnight activity, has demonstrated strong performance in detecting sleep and wake patterns in individuals with and without sleep disorders [14]–[16]. This suggests that actigraphy could pragmatically monitor sleep issues associated with TBI.

This project aims to determine the optimal time window after trauma for accurately classifying TBI by analyzing sleep and wake variables. Effective wearable and digital biomarkers were identified that could differentiate between individuals with and without TBI. By investigating how these biomarkers evolve over time and their correlation with sleep patterns, the project seeks to enhance diagnostic accuracy and improve TBI management through advanced monitoring methods. To fully leverage the temporal dynamics of these biomarkers, computational approaches capable of capturing subtle, time-dependent patterns are required. Given the complexity and variability of physiological responses following trauma, advanced machine learning techniques offer promising tools for this task. Numerous studies have leveraged artificial neural networks (ANNs) and machine learning (ML) models to predict TBI outcomes and severity, showcasing significant improvements in diagnostic accuracy, sensitivity, and specificity compared to traditional methods [17]–[25]. In this project, we take it a step further by utilizing a deep learning model with an attention mechanism that can selectively focus on the most informative time periods within the data. This allows the model to dynamically weigh the relevance of different time segments based on how far they occur from the traumatic event and how long the data was collected. By doing so, we aim to uncover critical temporal patterns that differentiate TBI+ from TBI− patients and identify optimal monitoring durations that maximize classification performance.

II. MATERIALS AND METHODS

A. Dataset and Preprocessing

In this project, we are utilizing the Advancing Understanding of RecOvery afteR traumA (AURORA) study dataset¹, which includes people who presented to the emergency department (ED) within 72 hours of a traumatic event. These individuals, aged 18 to 75, were assessed for eligibility to participate in the study based on specific qualifying traumatic events. Enrollment took place at participating sites for those meeting the criteria, which include motor vehicle collisions, physical or sexual assaults, falls from heights greater than 10 feet, and mass casualty incidents. Participants were required to wear a research wristwatch (Verily Life Sciences, San Francisco) for a minimum of 21 hours each day during the study period (see Fig. 1). However, the exact amount of time the wristwatch was worn differed depending on each participant’s circumstances. The methodology of the AURORA study, which was reviewed and approved by the Institutional Review Board (IRB #17–0703) at UNC Chapel Hill, has been previously published [26].

Fig. 1.

Schematic of the data collection process. Participants aged 18 to 75, admitted to the emergency department within 72 hours of specific traumatic events, were enrolled automatically in the study. They were required to wear a research wristwatch (Verily Life Sciences) for at least 21 hours daily for the next couple of weeks.

The custom-designed wristwatch used in this study was capable of capturing a broad spectrum of physiological and behavioral data. It recorded photoplethysmogram (PPG) signals, which are valuable for assessing heart rate variability and overall cardiovascular health, as well as high-resolution motion data from a built-in tri-axial accelerometer. This sensor tracked movement along three perpendicular axes (X, Y, and Z) with a 12-bit resolution and a dynamic range of ±8g, offering a fine sensitivity of about 1 milligravity (0.001 g). The sampling rate was typically configured to 32 Hz, though this could be adjusted depending on the study needs. The device’s firmware ensured consistent timestamping and supported onboard storage, enabling extended data collection without frequent interruptions for charging or data transfer. Although not a consumer device, the wristwatch was engineered to meet rigorous research standards, delivering stable and accurate data over prolonged monitoring durations.

Activity counts were generated from the accelerometer data collected at around 30–32 Hz by identifying the peak acceleration values within each one-second window and summing these over consecutive 30-second intervals to quantify physical activity levels throughout the day [27]. The raw data were subsequently processed using the MATLAB-based Actigraphy Toolbox to compute actigraphy counts [28]. Figure 2 represents a ”double plot” visualization of actigraphy data collected over a 28-day period (x-axis) for four different participants. Each main heatmap panel displays 48 hours of data per column—e.g., the first column represents days 1 and 2, the second days 2 and 3, and so on—enabling clearer visualization of repeating daily activity and rest patterns. To the right of each heatmap, an additional panel depicts the average activity for each hour of the day, computed across all columns (i.e., across days). This provides a compact summary of the underlying circadian rhythm, highlighting periodic patterns in rest and activity over the full observation period.

Fig. 2.

Double plot displaying actigraphy data for four patients, with each plot illustrating movement intensity over time. The x-axis represents the first 28 days post-trauma, and the y-axis represents 48 hours per column—e.g., each column overlays two consecutive days (e.g., days 1–2, 2–3, etc.)—to highlight daily and circadian patterns. Color gradients indicate movement intensity, with lighter colors corresponding to higher activity levels. To the right of each heatmap, an additional panel shows the average activity for each hour across all days, providing a compact summary of circadian rhythm patterns over the 28-day period.

At the time of trauma, if there is mechanical damage to the brain, Glial Fibrillary Acidic Protein (GFAP)—an intermediate filament protein predominantly found in astrocytes, the most common type of cell in the central nervous system—can increase in the bloodstream. Elevated peripheral levels of GFAP were approved by the U.S. FDA in 2018 [29] as a key biomarker to assist clinicians in ruling out the need for CT scanning in mild traumatic brain injury cases, rather than for directly detecting TBI [30], [31]. However, emerging evidence suggests that GFAP could play a pivotal role in identifying TBI more broadly. Notably, Yue et al. (2019) [32] demonstrated a compelling association between GFAP levels and TBI severity, providing a strong rationale for its use as a diagnostic biomarker in TBI detection. Determining the right cutoff value for GFAP to identify TBI can vary widely depending on the study and the criteria used. In this project, we assessed three different cutoff values to see how well they distinguish between TBI and non-TBI patients [29]. To determine our primary threshold for identifying TBI using GFAP levels, we referred to the largest study reporting GFAP concentrations in healthy individuals [33]. This study, which used the i-STAT system², reported a median GFAP concentration of 10 pg/mL, with an interquartile range (IQR) of 5.0–15.0 pg/mL. We used Tukey’s method for outlier detection³, which identified values above 30 pg/mL as outliers. Since GFAP measurements differ across platforms, we converted this i-STAT cut-off to equivalent values on the Quanterix platform⁴. Calibration data indicated that 30 pg/mL on the i-STAT corresponds to approximately 268 pg/mL on Quanterix. We thus used 268 pg/mL as our primary GFAP threshold for TBI identification. We also examined more stringent alternatives, including a threshold for extreme outliers (GFAP > 360 pg/mL) and a level suggested by a meta-analysis, which proposed a 100 pg/mL cut-off in i-STAT units (equivalent to 695 pg/mL on Quanterix) [34]. Therefore, the least conservative cutoff value is $268\frac{pg}{ml}$. This means it’s more sensitive and could identify more potential TBI cases, but it might also lead to more false positives—where people without TBI are incorrectly identified as having it. The second cutoff value is 360 $\frac{pg}{ml}$, which offers a balance between sensitivity and specificity, aiming to cut down on false positives while still capturing a good number of true TBI cases. Lastly, the third cutoff value is 695 $\frac{pg}{ml}$, the most conservative of the three. This higher threshold is intended to increase specificity, minimizing false positives as much as possible. While this approach lowers the chance of incorrectly diagnosing TBI in people who don’t have it, it might also miss some true TBI cases, resulting in lower sensitivity.

B. Feature extraction - Low-resolution features

Low-resolution features, which are daily-based, are extracted from the accelerometer data collected from patients’ wristwatches. These features include Total Sleep Time, representing the total minutes the patient sleeps (i.e. blue region) during the main sleep window (i.e. onset to offset region) in Fig. 3. Wake Minutes refer to the total minutes the patient is awake during the main sleep window, which is shown with the yellow and green regions between the sleep onset and offset in Fig. 3. Sleep Fragmentation is measured by the number of awakenings during the main sleep window, corresponding to the number of yellow and green regions in Fig. 3. The Longest Wake/Sleep Episodes are represented by the total minutes of the longest wake or sleep period within the main sleep window, highlighted by the largest yellow or blue region in the main sleep window. Mean Sleep Episodes is calculated as the average number of minutes the patient slept between awakenings, derived by dividing the total sleep time by the number of blue regions in Fig. 3. Mean Wake Episodes, on the other hand, represent the average minutes of wake periods, obtained by dividing the total minutes spent awake by the total number of wake periods; this is illustrated by dividing the total minutes of the yellow regions within the main sleep period by the number of yellow regions in Fig. 3. Sleep Period Duration reflects the total minutes the patient remained lying down during the main sleep window, excluding the green regions between the sleep onset and offset in Fig. 3. Sleep Percentage is calculated by dividing the total sleep minutes (i.e. blue region between the sleep onset and offset) by the time the patient remained lying down during this period, including both blue and yellow regions between the sleep onset and offset in Fig. 3, and then multiplied by 100%. Finally, the Average Euclidean Norm Minus One (ENMO) is calculated in 5-second intervals over 24 hours. This metric adjusts the raw 3-dimensional accelerometer data by subtracting 1g, the expected gravitational acceleration, from the magnitude of the total acceleration vector, setting any negative values to zero to ensure that only movement-related acceleration is captured while minimizing the influence of gravity.

Fig. 3.

The schematic illustrates different stages of the sleeping process. The yellow region denotes periods when the person is awake but lying down. The green region indicates times when the person is either visiting the restroom during the night or not lying down. The blue region represents periods when the person is asleep.

Before cleaning, the dataset included 501,630 patient days from 3,360 patients. The cleaning process involved several steps: first, we excluded patient days without sleep data, removing 113,212 days. Next, we eliminated data from 117 participants who withdrew from the study, accounting for 4,441 patient days. Additionally, we excluded days where the Total Sleep Time was 840 minutes or more and those where the Longest Sleep Episode was 5 minutes or less, removing another 4,760 days. After these steps, the dataset was reduced to 379,217 patient days from 3,216 patients. Ultimately, the number of patients was further reduced to 2,305. This reduction occurred because low-resolution features could not be extracted for all subjects in the initial dataset, likely due to factors such as low signal quality index or other data quality issues. As a result, we only retained patients for whom complete data was available across all three cut-off values.

C. Feature extraction - High-resolution features

High-resolution features are minute-based and include the activity count, which is derived from accelerometer measurements taken at a rate of 30 Hz to capture movement intensity throughout the day. This process involves aggregating peak values observed in one-second intervals and summing them for each 30-second segment of data [27]. To reduce computational cost, we opted for a 5-minute time window instead of the original 30-second window. Within each 5-minute interval, we averaged the extracted features from each 30-second segment. This approach effectively reduced the total number of time epochs.

To maintain consistency with the low-resolution features, we utilized data from the first 7 days, 14 days, and 21 days. For handling missing data, we employed bootstrapping based on the total available data [35]. To prepare the data for deep learning, we applied zero padding to ensure uniformity in data size across all datasets.

D. Bivariate analysis - Low-resolution features

In our analysis of low-resolution features, we computed average values for three different time windows: the first 7 days, 14 days, and 21 days. We then conducted t-tests (assuming data coming from a normal distribution) and Mann-Whitney tests (assuming data are not coming from a normal distribution) to compare these averages between TBI+ and TBI− groups, defining by cutoff values of $268\frac{pg}{ml}$, $360\frac{pg}{ml}$ and $695\frac{pg}{ml}$ on the GFAP level.

E. Deep learning model - High-resolution features

In our analysis of high-resolution features, we utilized activity count data from the entire dataset, which includes 2305 patients, recorded every 5 minutes. We employed a deep neural network model combining Residual blocks and a temporal attention mechanism to identify and prioritize the most critical parts of the input data. The model, as shown in Fig. 4 – Panel A, begins with an input layer receiving sequences of shape (T, 2), followed by a masking layer to handle the zero-padded data used to standardize input sizes. Three Residual blocks are stacked sequentially with filter sizes of 32, 128, and 32, respectively. Each block contains two 1D convolutional layers (kernel size = 3) with batch normalization and includes a skip connection via a 1 × 1 convolution to preserve gradient flow and address the vanishing gradient problem. A temporal self-attention layer enhances the model’s ability to focus on the most relevant segments of the time series data, ensuring that critical input is not treated uniformly. The network concludes with a global average pooling layer, a dropout layer (rate = 0.4), a dense layer with 128 ReLU-activated units, and a final softmax output layer⁵.

Fig. 4.

Panel A - Neural Network Architecture featuring Masking, Residual Blocks, and an Attention Mechanism. Panel B - Attention weights plotted for all patients, with time steps/epochs on the x-axis. Panel C - A zoomed-in view of Panel B, focusing on a subset of patients on the y-axis.

Model training was performed using the Adam optimizer (learning rate = 0.001), categorical cross-entropy loss, and early stopping (patience = 5). We used stratified 5-fold cross-validation to improve generalization, and addressed class imbalance via computed class weights. Prior to training, the second input feature (BootStrapFeatures) was standardized using a StandardScaler. Figure 4 – Panel B presents a heatmap of attention weights throughout the training process, with the y-axis showing the weights for different patients and the x-axis representing various time steps or epochs. The time steps are calculated using the formula: total time × 24 (hours) × 60 (minutes) ÷ 5 (minutes). As mentioned previously, the colors in the heatmap indicate attention weights, with lighter shades representing higher values and darker shades showing lower values. Panel C offers a zoomed-in view of Panel B, concentrating on a subset of patients along the y-axis.

III. RESULTS

A. Low-resolution data analysis

Among the ten features analyzed, only three (i.e. Average ENMO, Mean Wake Episodes and Longest Wake Episodes) exhibited significant differences between the TBI+ and TBI− classes for the first 7 days of data. According to Table I, as the number of days in the time window increases, the p-value from the t-test - under the assumption of normality—also increases. This trend is particularly evident for features like the longest wake episode, where no significant difference in average values is observed between the TBI+ and TBI− classes, regardless of the cutoff chosen. We also applied the Mann-Whitney test to account for the possibility that the data do not follow a normal distribution. The results were consistent with the t-test, supporting that early post-trauma data provide more informative insights. While the p-value increased as the time window length extended from 7 to 21 days across all three feature values, the average ENMO exhibited greater robustness to this change. This suggests that average ENMO may be a strong indicator for distinguishing between TBI+ and TBI− classifications with low-resolution data.

TABLE I.

Impact of time window length (7 to 21 days) on the significance levels of low-resolution features for TBI+ and TBI−, under the null hypothesis that TBI+ and TBI− classes are drawn from the same distribution.

		Average ENMO		Mean wake episodes		Longest wake episodes
GFAP Cut-off	Time Window	T-test	Mann-Whitney	T-test	Mann-Whitney	T-test	Mann-Whitney
	7-days	≤ 0.001***	≤ 0.001***	≤ 0.01**	≤ 0.001***	≤ 0.01**	≤ 0.01**
$\mathbf{268}\frac{pg}{ml}$	14-days	≤ 0.001***	≤ 0.001***	≤ 0.01**	≤ 0.01**	≤ 0.05***	≤ 0.01**
	21-days	≤ 0.001***	≤ 0.001***	≤ 0.05***	≤ 0.05***	0.063	≤ 0.05***
	7-days	≤ 0.001***	≤ 0.001***	≤ 0.01**	≤ 0.001***	≤ 0.01**	≤ 0.001***
$\mathbf{360}\frac{pg}{ml}$	14-days	≤ 0.001***	≤ 0.001***	≤ 0.01**	≤ 0.01**	≤ 0.05***	≤ 0.01**
	21-days	≤ 0.001***	≤ 0.001***	≤ 0.05***	≤ 0.01**	0.080	≤ 0.05***
	7-days	≤ 0.001***	≤ 0.001***	≤ 0.05***	≤ 0.01**	0.101	≤ 0.05***
$\mathbf{695}\frac{pg}{ml}$	14-days	≤ 0.001***	≤ 0.001***	0.054	≤ 0.05***	0.255	0.082
	21-days	≤ 0.001***	≤ 0.001***	0.088	0.051	0.267	0.163

Significance thresholds are denoted as ***p ≤ 0.001, **p ≤ 0.01, *p ≤ 0.05. Non-significant values (p > 0.05) are highlighted in red.

B. High-resolution data analysis

The input data for the deep learning model (see Fig. 4 - Panel A) is categorized into TBI+ and TBI− cases using the least conservative GFAP cutoff value (i.e., $268\frac{pg}{ml}$). To ensure a more accurate and generalizable evaluation of model performance, we employed a 5-fold cross-validation approach. To investigate temporal patterns in model interpretability, we averaged the attention weights across patients and across all folds using the $268\frac{pg}{ml}$ GFAP cutoff and a 21-day post-injury time window. The resulting mean attention weights over time are presented in Fig. 5. As shown, attention is highest during the first 7 days following the injury and gradually decreases over the subsequent two weeks. This trend suggests that data collected earlier-closer to the time of injury—carries greater importance for distinguishing between TBI-positive and TBI-negative cases.

Fig. 5.

Mean attention weights over time across TBI cases using the $268\frac{pg}{ml}$ GFAP cutoff. The plot shows the average attention weights from the RNN model over a 21-day post-injury period (This analysis corresponds to the shaded region in Table II, which highlights results for the same GFAP cutoff value and time window.), aggregated across all cross-validation folds. Horizontal dashed lines indicate the mean attention weight within each 7-day interval: days 0–7 (red), days 8–14 (yellow), and days 15–21 (orange). The elevated attention during the first week suggests that early post-injury data contribute most significantly to the model’s classification decisions.

Motivated by this observation, we investigated whether limiting input data to the first 7 days post-injury would improve the classification performance of the deep learning model, compared to using data from the first 14 or 21 days. Using the $268\frac{pg}{ml}$ GFAP cutoff, the RNN achieved sensitivity, specificity, and F1 scores of 81.1%, 24.9%, and 22.5% for the 7-day window; 39.9%, 66.0%, and 20.7% for the 14-day window; and 44.8%, 58.1%, and 20.1% for the 21-day window (see Table II). These results support the hypothesis that data collected closer to the injury event is more informative for distinguishing between TBI-positive and TBI-negative cases.

TABLE II.

Comparison of sensitivity, specificity, and F1 score across three GFAP cut-off values and classification methods (RNN, Logistic Regression, and SVM) for time windows of 7, 14, and 21 days.

GFAP Cut-off	Time Window	RNN			LogReg			SVM
		Sens	Spec	F1	Sens	Spec	F1	Sens	Spec	F1
	7-days	0.811	0.249	0.225	0.160	0.858	0.147	0.228	0.879	0.217
$\mathbf{268}\frac{pg}{ml}$	14-days	0.399	0.660	0.207	0.121	0.865	0.116	0.146	0.920	0.169
	21-days	0.448	0.581	0.201	0.082	0.889	0.087	0.110	0.937	0.141
	7-days	0.722	0.411	0.189	0.105	0.895	0.097	0.158	0.909	0.153
$\mathbf{360}\frac{pg}{ml}$	14-days	0.679	0.350	0.166	0.081	0.920	0.086	0.077	0.953	0.099
	21-days	0.866	0.182	0.172	0.067	0.921	0.072	0.057	0.964	0.081
	7-days	0.868	0.278	0.116	0.041	0.947	0.041	0.016	0.980	0.024
$\mathbf{695}\frac{pg}{ml}$	14-days	0.496	0.493	0.093	0.016	0.965	0.020	0.000	0.992	0.000
	21-days	0.454	0.575	0.099	0.016	0.975	0.022	0.000	0.993	0.000

When compared to the RNN, both Logistic Regression (LogReg) and Support Vector Machine (SVM) models consistently underperformed in terms of F1 score across all time windows. For the 7-day window, LogReg and SVM achieved F1 scores of 14.7% and 21.7%, respectively—both lower than the RNN’s 22.5%. The performance gap widened over longer durations: at 14 days, LogReg and SVM yielded F1 scores of 11.6% and 16.9%, respectively, versus 20.7% for the RNN; and at 21 days, their F1 scores dropped to 11.0% and 14.1%, compared to 20.1% for the RNN. While LogReg and SVM exhibited higher specificity (e.g., up to 94.1% for SVM at 21 days), they suffered from extremely low sensitivity (as low as 8.2% for LogReg and 10.1% for SVM at 21 days), which severely limited their overall classification effectiveness. In contrast, the RNN provided a more balanced trade-off between sensitivity and specificity, making it a more suitable approach for early TBI classification using time-dependent data.

IV. DISCUSSION

In our study, we utilized the largest dataset to date, encompassing 501,630 patient days from 3,360 individuals, and covering the longest observation period. It is also the first study to apply an attention model to identify the optimal post-trauma time window for classifying TBI using sleep/wake features. Our analysis reveals that the deep learning model assigns significantly higher attention to data from the first 7 days after injury, with a marked decrease in attention in the following weeks. These findings underscore the critical importance of early post-injury data for distinguishing between TBI-positive (TBI+) and TBI-negative (TBI−) cases. Specifically, data from the first 7 days post-injury achieved a sensitivity of 61%, specificity of 67%, and an F1 score of 38%. As the data collection period extended to 14 days (sensitivity: 51%, specificity: 68%, F1 score: 33%) and 21 days (sensitivity: 100%, specificity: 7%, F1 score: 30%), the model’s performance declined, highlighting the diminishing value of later data. By identifying the most informative periods for sleep/wake data, this study aims to enhance TBI classification accuracy and reliability, emphasizing the potential for early and accurate detection to improve patient outcomes through timely, targeted interventions.

The early and accurate diagnosis of TBI has been widely recognized as a critical factor in improving patient outcomes, as highlighted in several clinical studies [36], [37]. Advances in machine learning and deep learning have significantly contributed to our understanding and diagnosis of TBI, leveraging diverse datasets such as neuroimaging data [38]–[41] and audio-based detection techniques [42]–[44].

While some of these studies have incorporated transformer-based models for audio [45] and video data [46], to the best of our knowledge, this research introduces a novel approach by employing an attention mechanism to identify the optimal time window following trauma. This approach is designed to enhance the classification accuracy in distinguishing TBI-positive cases from non-TBI cases, addressing a critical gap in current methodologies. An important consideration in deploying machine learning models in clinical settings is their ability to remain robust in the face of data drift—that is, changes in the underlying data distribution over time due to evolving patient behaviors, clinical practices, or measurement conditions. This poses a challenge for models trained on historical data, which may not fully capture future variability. To address this, our framework incorporates an attention mechanism, which enables the model to dynamically prioritize the most informative portions of the input sequence. Rather than assigning equal importance to all time points, the attention block learns to emphasize contextually relevant segments, thereby enhancing the model’s adaptability to distributional changes. This temporal flexibility improves robustness by allowing the model to focus on clinically meaningful patterns that may emerge differently over time, offering a degree of resilience against moderate shifts in data distribution.

One potential limitation of our study is the possibility of inconsistent wearing of wrist-worn devices, which can lead to inaccurate activity detection. While we did not have access to direct wear compliance data to explicitly identify or correct for such instances, our feature extraction strategy was designed to mitigate this risk. In particular, we incorporated both low-resolution (daily-based) and high-resolution (minute-based) features that emphasize broader behavioral patterns rather than transient signal variations. The low-resolution features—such as Total Sleep Time, Sleep Fragmentation, Sleep Percentage, and Sleep Period Duration—capture aggregated measures over the nightly sleep window, making them less sensitive to brief device misplacement or dropout. Similarly, high-resolution features, which reflect movement intensity over 5-minute intervals, were derived using averaging techniques that reduce the influence of short-term irregularities. This dual-scale approach enhances the model’s resilience to noise introduced by suboptimal device usage. Furthermore, we applied data quality screening to exclude prolonged periods of inactivity occurring outside expected sleep windows, which could suggest device non-wear. Although the absence of objective compliance data remains a limitation, the robustness of our feature design likely helped minimize the impact of inconsistent device use on model performance.

Future research should validate and refine these findings by incorporating additional features, such as heart rate variability and multiscale entropy data, alongside actigraphy. This comprehensive approach may further boost the F1 score and other key performance metrics, increasing sensitivity and specificity in the early stages of detection. Additionally, quantitatively correlating proportions of sleep architecture—rapid eye movement (REM) sleep and non-REM, slow wave sleep disturbances—and incorporating gait impairment analysis via 3D accelerometry could further strengthen this approach. Subtle abnormalities in gait [47] and significant disruptions in both REM sleep, slow wave sleep, and autonomic tone could potentially extend the time accuracy for distinguishing between TBI+ and TBI− cases, ultimately leading to more effective diagnosis and treatment strategies.

V. CONCLUSION

These findings emphasize the crucial role of early monitoring and data collection in the days immediately following trauma. Sleep duration data provides the most valuable insights for accurate TBI classification when collected during the first week post-trauma. The model demonstrates that the first week is critical for identifying TBI cases. This study highlights the potential of utilizing temporal data patterns in medical diagnostics, particularly for TBI. We can significantly improve diagnostic models and enhance patient care by concentrating on the early post-trauma period.