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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Closed-loop phase-locked transcranial magnetic stimulation (TMS) requires precise real-time electroencephalography (EEG) phase prediction.
• Timing inaccuracies, especially near EEG signal peaks and troughs, significantly impair targeting accuracy.
• Existing prediction methods face challenges in achieving the low latency and high consistency needed for effective closed-loop control.

Purpose of the Study:

• To benchmark classical and recurrent neural network (RNN) predictors for EEG phase prediction.
• To develop novel deep learning models that enhance phase prediction consistency and reduce timing lag, particularly around signal extrema.
• To introduce a new metric, Mean Lag Time (MLT), for evaluating extremum-specific prediction performance.

Main Methods:

• Benchmarking AR, FFT, LSTM, and GRU predictors on the Monash University TEPs-MEPs dataset.
• Proposing a parallel DSC-Attention-GRU architecture for efficient spatio-temporal feature extraction and attention-based dependency modeling.
• Developing a lightweight SqueezeNet-Attention-GRU variant for real-time applications and evaluating performance using MLT, PLV, APE, MAE, and RMSE.

Main Results:

• LSTM and GRU models showed improved temporal dynamics over AR/FFT but retained residual lag.
• The proposed DSC-Attention-GRU model consistently improved phase prediction accuracy and reduced extremum lag (MLT decreased from ~7.77-7.79 ms to ~7.50-7.56 ms).
• The lightweight variant achieved a 3.7% inference speedup while maintaining stable performance.

Conclusions:

• Explicitly optimizing extremum timing using MLT is crucial for closed-loop TMS.
• Integrating depthwise separable convolutions (DSC) and attention mechanisms enhances multi-channel modeling for reduced peak/trough lag.
• The developed models offer improved phase-consistent prediction, supporting the advancement of low-latency closed-loop phase-locked TMS.