Representation learning identifies systems-level neuroimaging signatures of traumatic brain injury-related attentional dysfunction in young adults

Background. Traumatic brain injury (TBI) is frequently followed by persistent deficits in attention and inhibitory control, with underlying neurobiological signatures poorly understood. Neuroimaging studies have identified widespread while inconsistent brain alterations after TBI, partly because TBI affects distributed brain systems rather than focal regions. We aimed to determine whether representation learning applied to multimodal, systems-level neuroimaging features could identify clinically interpretable signatures of TBI-related attentional dysfunction in young adults. Method. We studied 89 young adults (44 TBI, 45 controls). Muti-modal neuroimaging data were processed to derive 997 region-of-interest-based and systems-level features. After leakage-controlled feature reduction, a semi-supervised autoencoder was trained to learn compact latent neurobiological representations while reconstructing input features and classifying TBI status. Permutation-based feature importance and partial least squares regression were used to evaluate model interpretability and brain–behavior relationships. Results. An average area under receiver operating characteristic curve of 0.932 for group discrimination was achieved. Permutation analysis identified 6 highly discriminative systems-level features from the gray matter functional network and white matter structural network. In the TBI group, partial least squares regression showed that among these 6 features, nodal degree of left posterior superior temporal sulcus in the functional network , nodal global efficiency of right rostral anterior cingulate region in the structural network, and nodal global efficiency of right caudal anterior cingulate region in structural network explained substantial variance in inattentive and hyperactive/impulsive symptoms. Conclusions. Representation learning identified latent systems-level neuroimaging signatures that accurately differentiated chronic TBI from healthy control status and were meaningfully associated with attention-related symptom severity. The proposed framework provides an interpretable multimodal approach for developing biologically grounded biomarkers of post-TBI cognitive dysfunction.