Direct Prediction of the Complex-Valued Analytic Signal of EEG from Raw Multichannel Data

Accurate estimation of instantaneous neural dynamics is essential for electroencephalography (EEG)-based brain-state analysis and future closed-loop applications. Conventional phase estimation methods that rely on bandpass filtering and autoregressive prediction are limited by reduced accuracy near the current time point, poor future prediction capability, and susceptibility to inconsistencies in the contributing electrodes. This study proposes a deep neural network (DNN) framework that directly predicts the complex-valued analytic signal of EEG--representing instantaneous phase and amplitude--from raw multichannel EEG data. The model employs a temporal convolutional encoder and a probabilistic output head that jointly estimate the mean and variance of analytic signals at multiple past and future time points. To enhance robustness, a missing-signal imitation (MSI) mechanism is applied during training, together with electrode-mixing strategies incorporated into the model. Using resting-state EEG from 25 participants, we show that the proposed model consistently outperforms an autoregressive baseline in both subject-wise and cross-subject evaluations. MSI further improves stability when critical electrodes are removed and yields better calibration of predictive uncertainty. These findings demonstrate that analytic EEG signals can be estimated directly from raw data without manual preprocessing or subject-specific calibration. The proposed framework provides a scalable basis for automated EEG analysis and offers strong potential for future real-time and closed-loop neural applications.