Interpretable Transformer Models for rs-fMRI Epilepsy Classification and Biomarker Discovery

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Automated interpretation of resting-state fMRI (rs-fMRI) for epilepsy diagnosis is challenging.
• Developing robust methods for classifying epilepsy using rs-fMRI is crucial for improving patient outcomes.

Purpose of the Study:

• To develop and evaluate a regularized transformer model for epilepsy classification using rs-fMRI.
• To identify interpretable, network-level candidate biomarkers for epilepsy.

Main Methods:

• Utilized Schaefer-200 parcel time series from rs-fMRI data preprocessed with fMRIPrep.
• Developed a regularized transformer model incorporating attention mechanisms, learned positional encoding, and fMRI-specific regularization.
• Trained and validated the model using 4-fold cross-validation on a cohort of 65 participants (30 epilepsy, 35 controls) and an independent external dataset.

Main Results:

• The regularized transformer achieved high performance in within-fold classification (e.g., Accuracy 0.77, AUC 0.76).
• External validation showed promising but lower performance (Accuracy 0.60, AUC 0.64).
• Attribution-guided analysis identified candidate biomarkers in limbic, somatomotor, default-mode, and salience networks.

Conclusions:

• A regularized transformer model shows potential for classifying epilepsy from rs-fMRI and generating interpretable biomarkers.
• Preliminary results suggest the model's efficacy, but larger multi-site validation and stability testing are necessary for clinical translation.