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

Technical Summary: Representation Learning Identifies Systems-Level Neuroimaging Signatures of TBI-Related Attentional Dysfunction

Problem Statement
Traumatic brain injury (TBI), particularly mild-to-moderate cases, frequently results in persistent attentional deficits and inhibitory control impairments that significantly impact quality of life. However, the underlying neurobiological signatures remain poorly understood. Traditional neuroimaging studies have yielded inconsistent findings regarding specific brain regions or pathways involved in TBI-related attentional dysfunction. This inconsistency arises partly because TBI often involves diffuse neuropathological processes affecting distributed neural systems rather than focal lesions, which conventional voxel-wise or region-of-interest (ROI) analyses fail to capture effectively. Furthermore, existing machine learning (ML) approaches in TBI research have largely focused on diagnostic group discrimination using small datasets and supervised classifiers, with limited emphasis on identifying neurobiological features specifically linked to attentional symptom severity or utilizing the latent structure of high-dimensional multimodal data.

Methodology
The study employed a semi-supervised representation learning framework applied to multimodal neuroimaging data from 89 young adults (44 with a history of sports-related TBI and 45 healthy controls).

• Data Acquisition and Processing: Participants underwent high-resolution structural MRI (sMRI), diffusion MRI (dMRI), and task-based functional MRI (fMRI) using a validated visual sustained attention task (VSAT).

  • sMRI: Processed via the Human Connectome Project pipeline and FreeSurfer to extract 157 features (cortical thickness, volume) from 89 ROIs.
  • dMRI: Preprocessed using FSL to extract 384 features, including fractional anisotropy (FA) from 14 white matter tracts and graph-theoretical metrics (nodal global/local efficiency, clustering coefficient, betweenness centrality) for the white matter structural network.
  • fMRI: Processed using the GAT-FD toolbox to construct functional networks from 114 task-activated ROIs, yielding 456 features (nodal degree, efficiency, betweenness centrality).
  • Total Features: 997 raw multimodal features were generated per subject.

• Feature Reduction: To mitigate overfitting and reduce dimensionality, a leakage-controlled, two-stage feature selection procedure was applied within each cross-validation fold. This involved two-sample t-tests ($p < 0.05$) followed by Linear Discriminant Analysis (LDA) to retain features with absolute coefficient scores $>1.5$, reducing the feature space from 997 to 29.

• Model Architecture: A semi-supervised autoencoder (AE) was implemented with three components:

  1. Encoder: Compressed the 29-dimensional input into a 10-dimensional latent representation via a hidden layer of 20 neurons.
  2. Decoder: Reconstructed the original feature space using a mirrored architecture.
  3. Classifier: A supervised branch that predicted TBI status using the latent representation (5 neurons, sigmoid output).
     The model jointly optimized mean squared reconstruction error (weight 0.7) and binary cross-entropy classification loss (weight 0.3). Gaussian noise injection (20% of inputs) served as a denoising regularization strategy.

• Evaluation and Interpretability:

  • Classification: Performance was evaluated via 5-fold cross-validation using AUC, accuracy, sensitivity, and specificity.
  • Feature Importance: Permutation-based analysis (1,000 shuffles) identified highly discriminative features (threshold: mean + 1.5 SD).
  • Brain-Behavior Analysis: Partial Least Squares Regression (PLSR) was performed within the TBI group to assess the relationship between the top 6 discriminative features and attentional symptoms (measured by CAARS T-scores), controlling for age, sex, injury severity, and number of injuries.

Key Results

• Classification Performance: The semi-supervised AE model achieved an average Area Under the Curve (AUC) of 0.932 ± 0.087, with an accuracy of 88.5%, sensitivity of ~90%, and specificity of ~86%. This performance significantly outperformed a conventional PCA-SVM benchmark.
• Discriminative Features: Permutation analysis identified a small subset of highly discriminative systems-level features. The top 4 features (exceeding mean + 1.5 SD) were:
  1. Nodal betweenness centrality (BC) of the right caudate nucleus (functional network).
  2. Nodal degree of the right cingulate gyrus (functional network).
  3. Nodal degree of the left posterior superior temporal sulcus (pSTS) (functional network).
  4. Nodal BC of the right cuneus (white matter structural network).
     Two additional features (nodal global efficiency of right rostral and caudal anterior cingulate regions in the structural network) exceeded the mean + 0.5 SD threshold.
• Feature Ablation: A nonlinear relationship was observed between feature count and performance; the top 4 features captured the majority of discriminative power, with the full 29-feature model providing marginal gains in stability.
• Brain-Behavior Relationships: In the TBI group, PLSR revealed that the identified features explained 65.7% of the variance in attentional symptoms ($R^2Y = 0.657$) with robust predictive relevance ($Q^2 = 0.534$, $p < 0.001$).
  • Left pSTS (Functional): Lower nodal degree was significantly associated with higher inattentive and hyperactive/impulsive symptoms (VIP = 1.48).
  • Right Rostral/Caudal Anterior Cingulate (Structural): Higher nodal global efficiency was positively associated with symptom severity (VIP > 1.0).
  • No significant brain-behavior associations were found in the control group.

Key Contributions

1. Framework Development: The study successfully applied a semi-supervised autoencoder to multimodal neuroimaging data, demonstrating that representation learning can effectively compress high-dimensional, heterogeneous TBI data into compact, biologically meaningful latent representations.
2. Systems-Level Signatures: Unlike previous studies focusing on focal regions, this work identified that TBI-related attentional dysfunction is characterized by distributed alterations in network properties (nodal degree, efficiency, betweenness centrality) across functional and structural networks.
3. Clinical Relevance: The study moved beyond simple diagnostic classification by demonstrating that the latent features learned by the model are significantly associated with the severity of attentional symptoms (inattentiveness and hyperactivity/impulsivity) within the TBI population.
4. Interpretability: Through permutation importance and PLSR, the study provided an interpretable mapping between specific network hubs (pSTS, cingulate, caudate, cuneus) and clinical outcomes, offering a biologically grounded basis for biomarker development.

Significance and Claims
The authors claim that this study provides a novel, interpretable multimodal approach for developing biologically grounded biomarkers of post-TBI cognitive dysfunction. By shifting the focus from isolated regional abnormalities to latent systems-level signatures, the framework captures the distributed nature of TBI pathology more effectively than conventional statistical or supervised ML methods.

The findings suggest that persistent attentional dysfunction after TBI is not merely a result of focal damage but reflects disrupted communication and organization across interacting neural systems, specifically involving fronto-striatal, cingulate, posterior temporal, and occipital networks. The authors posit that these latent neurobiological representations offer a promising avenue for objective stratification of cognitive dysfunction and may guide the development of targeted interventions. However, the authors modestly note that the current framework is not yet ready for immediate clinical implementation and requires validation in larger, multi-site cohorts with longitudinal follow-up to establish generalizability and causal mechanisms.