Lightweight EEG Phase Prediction Based on Channel Attention and Spatio-Temporal Parallel Processing

<b>Background/Objectives:</b> Closed-loop phase-locked TMS aims to deliver stimulation at targeted EEG phases, but real-time phase prediction remains a practical bottleneck. Timing errors are especially harmful near peaks and troughs, where small offsets can substantially degrade phase targeting. We benchmark representative predictors and develop models that improve phase consistency while reducing peak/trough lag. <b>Methods:</b> Using the publicly available Monash University TEPs-MEPs dataset, we benchmark classical predictors (AR- and FFT-based) and recurrent baselines (LSTM, GRU). To quantify extremum-specific behavior critical for closed-loop triggering, we propose Mean Lag Time (MLT), defined as the average temporal offset between predicted and ground-truth extrema, alongside PLV, APE, MAE, and RMSE. We further propose a parallel DSC-Attention-GRU architecture combining depthwise separable convolutions for efficient multi-channel spatio-temporal feature extraction with self-attention for spatial reweighting and dependency modeling, followed by a GRU phase predictor. A lightweight SqueezeNet-Attention-GRU variant is also designed for real-time constraints. <b>Results:</b> LSTM/GRU outperform AR/FFT in capturing temporal dynamics but retain residual peak/trough lag. Across stimulation intensities and frequency bands, DSC-Attention-GRU consistently improves phase consistency and prediction accuracy and reduces extremum lag, lowering MLT from ~7.77-7.79 ms to ~7.50-7.56 ms. The lightweight variant maintains stable performance with an average 3.7% inference speedup. <b>Conclusions:</b> Explicitly optimizing extremum timing via MLT and enhancing multi-channel modeling with DSC and attention reduces peak/trough lag and improves phase-consistent prediction, supporting low-latency closed-loop phase-locked TMS.