A Convolutional Neural Network Combining Multi-scale Convolution and Functional Connectivity for Predicting rTMS Efficacy in MUD.

PURPOSE: Repetitive transcranial magnetic stimulation (rTMS) is a promising neuromodulation approach for treating methamphetamine use disorder (MUD), but its therapeutic outcomes exhibit considerable variability. Therefore, this study proposed a deep learning model for predicting treatment response based on neuroscience findings, which may guide therapeutic decisions and prevent healthcare resource wastage.

METHODS

Resting-state electroencephalography (EEG) was acquired from 17 subjects with MUD prior to a 4-week rTMS treatment regimen. After treatment, subjects were classified as either responders or non-responders according to standardized clinical assessments. A convolutional neural network (MSFC-CNN) was developed that parallelly integrated multi-scale (MS) and functional connectivity (FC) modules. The MS module employed multi-scale kernels to extract temporal features at varying resolutions, enhanced by a frequency-domain self-attention mechanism that adaptively recalibrated spatiotemporal feature weights across different frequency bands. The FC module performed 2D convolution on functional connectivity matrices to leverage inter-channel dependencies and strengthen spatial representation learning. Predictions from both modules were fused at the decision level to generate the final classification.

RESULTS

MSFC-CNN achieved an average prediction accuracy of 88.06% [95% CI 80.36%, 95.75%] for classifying responders from non-responders. The proposed model statistically outperformed classical and state-of-the-art comparison methods (see "Results" for paired tests and effect sizes). Feature visualization and ablation studies further validated the contributions of each module within the network architecture. Additionally, the model also showed discriminative ability on the auxiliary MUD vs. HC task.

CONCLUSION

As an initial exploration, the proposed MSFC-CNN provides methodological support for predicting treatment responses in MUD based on EEG.