Exploiting Graph Convolutional Networks for Insightful Classification and Explanation of Traumatic Brain Injury

Understanding how brain connectivity is disrupted across stages of traumatic brain injury (TBI) is essential for improving diagnosis and treatment. TBI poses major challenges in clinical assessment, requiring advanced neuroimaging and machine learning (ML) for effective patient stratification. This study classifies TBI patients into acute, chronic, and control groups using graph convolutional networks (GCNs) applied to structural connectomes derived from diffusion-weighted imaging (DWI). To enhance interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was used to identify brain regions contributing to classification. The dataset included 40 participants: 18 acute TBI patients (Glasgow Coma Scale $\leq 6$ , enrolled after $\geq 24$ hours of unresponsiveness), 6 chronic patients with persistent disorders of consciousness, and 16 healthy controls. Nine acute patients who regained consciousness were later included in the chronic group to assess longitudinal changes. The GCN was trained on DWI-derived connectomes and evaluated using leave-one-subject-out (subject-wise) cross-validation. It achieved 83.67% accuracy, with precision, recall, and F1-score of 81.6%, 78%, and 79%, respectively, reported as per-fold averages. Grad-CAM identified thalamus, anterior cingulate cortex, and frontal cortex as key regions for group discrimination. Results suggest a shift from widespread neural disruption in acute TBI to more localized impairments in the chronic stage, possibly reflecting compensatory reorganization. Despite the limited sample size, the model's robustness was supported by conservative regularization and subject-wise validation. Notably, the GCN outperformed classical ML classifiers, offering superior accuracy and greater biological plausibility. These findings support the use of GCN-based pipelines for clinical decision support and highlight their potential to inform interpretable, ML-driven strategies for precision neurorehabilitation.