Enhanced EEG Forecasting: A Probabilistic Deep Learning Approach.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Accurate forecasting of electroencephalography (EEG) signals is crucial for real-time applications like brain-computer interfaces.
• Traditional autoregressive (AR) models have limitations in long-range EEG prediction accuracy.
• Advancements in deep learning offer potential for improved EEG signal forecasting.

Purpose of the Study:

• To enhance EEG signal forecasting using probabilistic deep learning.
• To compare the performance of the WaveNet model against the autoregressive (AR) model for EEG prediction.
• To investigate the utility of probabilistic deep learning for longer-range EEG forecasts.

Main Methods:

• Applied the probabilistic deep neural network model WaveNet to forecast resting-state EEG signals.
• Focused on theta (4-7.5 Hz) and alpha (8-13 Hz) frequency bands.
• Compared WaveNet's predictive accuracy, including signal amplitude and phase estimation, against the AR model.

Main Results:

• WaveNet reliably predicted EEG signals 150 ms ahead in theta and alpha bands with low mean absolute errors (1.0 ± 1.1 µV for theta, 0.9 ± 1.1 µV for alpha).
• WaveNet outperformed the AR model in predicting both signal amplitude and phase.
• The probabilistic approach enabled more accurate forecasting and effective discarding of uncertain predictions.

Conclusions:

• Probabilistic deep learning, specifically WaveNet, is effective for forecasting resting-state EEG time series.
• This approach offers superior accuracy and robustness compared to traditional AR models for EEG prediction.
• The developed model holds promise for enhancing real-time brain state estimation in BCI, brain stimulation, neuroscientific inquiry, and diagnostics.