Interpretable Multiscale Directed Temporal Graph Learning for MEG-Based Identification and Lateralization of Temporal Lobe Epilepsy

Abstract

Objective: To address the limitations of existing deep learning methods for brain networks in jointly modeling directional interregional connectivity, multiband information, and short-term dynamic features, this study proposes a multiscale directed temporal graph convolutional network (MSD-STGNN) for the three-class classification of healthy controls (HCs), patients with left temporal lobe epilepsy (lTLE), and patients with right temporal lobe epilepsy (rTLE). Methods: Resting-state magnetoencephalography (MEG) data were obtained from 43 subjects, including 14 HCs, 13 patients with lTLE, and 16 patients with rTLE. Based on 26 predefined default mode network (DMN)-related brain regions, directed effective connectivity networks were constructed in six frequency bands using the directed transfer function (DTF), and indices including information-flow strength, directional preference, and hemispheric asymmetry were used as node features. MSD-STGNN separately modeled incoming and outgoing connectivity information through directed graph convolution, fused frequency-band information using a hierarchical multiband attention mechanism with gated residual correction, and employed gated recurrent units (GRUs) to extract short-term dynamic features from consecutive brain-network slices. Model performance was evaluated using subject-level stratified fivefold cross-validation, with predictions from multiple temporal groups of each subject aggregated to obtain the final subject-level prediction. Frequency-band masking and node-level fusion-weight analyses were further performed to evaluate the model’s dependence on different frequency bands and information from the predefined brain regions. Results: In the subject-level fivefold cross-validation, MSD-STGNN achieved an accuracy of 0.836 ± 0.067, a macro-F1 of 0.830 ± 0.065, and a macro-AUC of 0.900 ± 0.055 using the one-vs-rest strategy, with the highest fivefold mean values across all evaluation metrics among the baseline models and ablation configurations investigated in this study. Post-training frequency-band masking showed that the model exhibited relatively high dependence on the low-gamma (30–80 Hz) and beta (13–30 Hz) bands. Node-level fusion-weight analysis showed that, within the 26 predefined DMN-related brain regions, orbitofrontal, cingulate, medial temporal, and parietal regions exhibited relatively high overall fusion weights across different frequency bands, although the exact top 5 regional rankings varied across folds. Significance: Within a unified framework, MSD-STGNN integrates directional connectivity, multiband information, and short-term dynamic features derived from MEG-based directed brain networks, providing a modeling approach with a certain degree of interpretability for the three-class classification of HCs, patients with lTLE, and patients with rTLE. The current findings are based on internal cross-validation of a small, single-center cohort; therefore, the classification performance and the observed frequency-band and brain-region attention patterns require further validation in larger, independent multicenter datasets.