University of Tabriz Tabriz Journal of Electrical Engineering 2008-7799 2026 05 30 Adaptive Masked Spatio-Temporal Graph Convolutions for Classification of Temporal Lobe Epilepsy from Resting-State Fmri Adaptive Masked Spatio-Temporal Graph Convolutions for Classification of Temporal Lobe Epilepsy from Resting-State Fmri 21655 10.22034/tjee.2026.71264.5118 FA Journal Article 2026 01 25 Temporal lobe epilepsy (TLE) is increasingly recognized as a network disorder involving widespread disruptions of functional brain connectivity. Resting-state functional MRI (rs-fMRI) captures these dynamics, but conventional methods often assume static connectivity or ignore inter-regional interactions. We propose an Adaptive Masked Spatio-Temporal Graph Convolutional Network (AdaMST-GCN), which learns sparse, data-driven adjacency masks from partial-correlation graphs and integrates them with temporal convolutions to model dynamic network patterns. Evaluated on a TLE rs-fMRI dataset using 5-fold cross validation with sliding windows (50, 100, 150, 200 TRs), AdaMST-GCN achieved a mean held-out test F1 score of 79.0%, outperforming the original ST-GCN (75.7%) and LSTM baseline (58.5%). At 50-TR windows, it peaked at 82.3% F1. The learned masks consistently identified high-centrality regions, including the precuneus, temporal pole, and orbitofrontal cortex, corresponding to known TLE pathology. These results demonstrate that adaptive graph learning improves both predictive accuracy and interpretability, providing clinically relevant biomarkers. Temporal lobe epilepsy (TLE) is increasingly recognized as a network disorder involving widespread disruptions of functional brain connectivity. Resting-state functional MRI (rs-fMRI) captures these dynamics, but conventional methods often assume static connectivity or ignore inter-regional interactions. We propose an Adaptive Masked Spatio-Temporal Graph Convolutional Network (AdaMST-GCN), which learns sparse, data-driven adjacency masks from partial-correlation graphs and integrates them with temporal convolutions to model dynamic network patterns. Evaluated on a TLE rs-fMRI dataset using 5-fold cross validation with sliding windows (50, 100, 150, 200 TRs), AdaMST-GCN achieved a mean held-out test F1 score of 79.0%, outperforming the original ST-GCN (75.7%) and LSTM baseline (58.5%). At 50-TR windows, it peaked at 82.3% F1. The learned masks consistently identified high-centrality regions, including the precuneus, temporal pole, and orbitofrontal cortex, corresponding to known TLE pathology. These results demonstrate that adaptive graph learning improves both predictive accuracy and interpretability, providing clinically relevant biomarkers.