(K-421) Do Individual Differences in Functional Connectivity Make Brain Trauma Unidentifiable? Hypothetically Exploring Optimal Brain States for Injury Detection

Traumatic brain injury (TBI), resulting from excessive external forces on the head, is recognized as a disconnection syndrome that rapidly disrupts brain networks, leading to behavioral and cognitive disorders. The evaluation of this mechanism was historically challenging until the emergence of functional magnetic resonance imaging (fMRI) and functional connectivity (FC), enabling the examination of intrinsic connectivity networks by assessing the functional correlation between different brain regions. fMRI studies detect activity changes during the “Resting state,” where subjects lie quietly in the scanner, or with specific tasks assigned. Group differences in resting state and task-based fMRI between TBI patients and healthy controls indicate alterations in functional connectivity. However, utilizing FC changes for clinical diagnosis and outcome prediction in individual TBI patients remains a significant challenge. Hypothetically, this study aims to identify the ideal fMRI condition for diagnosing TBI.

We first present observations on FC variability from the human connectome project (HCP) and examine how brain injury impacts FC using data from patients with acute severe TBI. We further investigate whether FC conditions are identifiable from the comprehensive HCP dataset of eight brain conditions through the implementation of a machine-learning model. To further understand the potential of fMRI in diagnosing TBI, we simulate virtually injured FC for each brain condition based on observations from patient images. Subsequently, we assess in which brain condition machine-learning models can best differentiate injurious cases from healthy subjects under varying injury severity scenarios.

This study included 992 healthy subjects from the HCP, providing complete data for eight scanning conditions: Resting state, emotion, gambling, language, motor, relational, social, and working memory. Additional Injury fMRI data with “language” stimulus were obtained from a public dataset comprising 8 subjects with acute severe TBI. Sixteen age and sex-matched control subjects were included for comparison. FC was processed and computed using the Conn toolbox and the Schaefer parcellation atlas containing 100 regions.

To capture FC changes efficiently, we employed “Sparse Connectivity Patterns” (SCPs), identifying the different presence (coefficients) of ten SCPs with a sparsity level of 25 as optimal representations. SCP coefficients effectively differentiated healthy control from TBI patients and variations between brain conditions. Building on this observation, we virtually injured FC by randomly altering the SCP coefficients with injury severity determined by the number and magnitude of changed coefficients. Virtually injured FC data were generated for each subject in the HCP database across various injury severity levels and conditions.

For FC classification, we employed the convolutional neural network. Initially, a deep learning model was optimized using pooled healthy HCP data accurately to identify the imaging conditions. Subsequently, we combined the virtually injured FC data with the original healthy FC data to optimize a machine-learning model capable of detecting injurious FC under each condition. The sensitivity and specificity of the model were assessed in relation to injury severity for each condition.

Pooling healthy FC data from different conditions, we observed that the “resting state” was the most successfully identified brain state, achieving a remarkable 100% accuracy using the machine learning model. In contrast, the “emotion" state posed the most significant challenge, with an accuracy of 92.9%.  This finding suggests that between-subject similarity during the resting state is higher than between-state similarity, indicating distinct patterns during different brain conditions. Additionally, there was a potential for confusion between the "emotion" states and other brain states, indicating overlapping FC patterns between these conditions.

Interestingly. When classifying injury status, the “resting state” was not consistently the best condition for differentiating virtual injury, depending on the severity of the injury. Instead, the “gambling” task displayed comparable, if not better, accuracy in identifying injury. The “emotion”, “social”, and “working memory” tasks also showed intermediate accuracy, while the “relational” task exhibited the lowest accuracy overall. These results suggest unique changes in functional connectivity during different tasks following brain injury.

Despite the valuable insights gained from this study, we must acknowledge certain limitations. First, we did not account for several other factors, such as acquisition and processing variation, which could influence FC results. Additionally, we did not explore the injury susceptivity of specific brain regions, which may impact injury patterns. The question of which brain condition is best for identifying injury may lack a universal answer, as it likely depends on the specific population and neurological deficit being studied. Nevertheless, this study investigated a substantial number of subjects scanned under diverse conditions following injury, offering crucial insights and proving that the brain state significantly influences injury detection.

In conclusion, our results reveal differential accuracy in identifying brain states, with the resting state being highly distinguishable and the "emotion" state presenting more challenges. Regarding injury detection, the optimal brain condition for differentiation varies based on injury severity, with the "gambling" and “resting states” task showing as potential ideal states. Despite limitations, this study underscores the impact of individual differences in functional connectivity on brain trauma identification and emphasizes the importance of considering specific brain states in TBI diagnosis.