Deep Neural Network for Direct Prediction of Analytic Signals from Neural Oscillations

Objective: A method for real-time prediction of analytic signals is needed for state-informed stimulation in electroencephalography (EEG) experiments. The currently available methods lack sufficient prediction accuracy or have a complicated selection process of experimental parameters. Approach: The proposed method uses a deep neural network (DNN) to predict current and future EEG phases and amplitudes from raw EEG data for real-time signal processing. Main results: The proposed method predicted EEG phases and amplitudes more accurately than the conventional method based on an autoregressive model for actual EEG data. Furthermore, the DNN incorporates a missing signal imitation layer to make the model robust against missing data due to recording failures in EEG experiments. Significance: The DNN allows the current EEG phase to be estimated directly from the raw EEG data, which reduces the number of experimental parameters required. The end-to-end prediction framework can help simplify experimental operations and facilitate the implementation of the proposed method in various applications concerning the modulation of neural dynamics.