Reconstruction of EEG signals using next-generation reservoir computing

European Physical Journal: Special Topics. 2026. P. 1–10.

Badarin A., Ratnikov F., Andreev A.

EEG recordings are often affected by the loss or corruption of individual channels due to electrode detachment, poor scalp contact, or external interference. Such channels must be accurately reconstructed before further analysis. In this study, we investigate Next-Generation Reservoir Computing (NG-RC) as a data-driven approach for reconstructing corrupted EEG channels and compare its performance with spherical spline interpolation implemented in MNE-Python. Using an 18-channel dataset from 24 healthy participants, we reconstruct each channel from the remaining channels and assess reconstruction quality in six frequency ranges: δ (1–4 Hz), θ (4–8 Hz), α (8–13 Hz), β (13–30 Hz), γ (30–40 Hz), and the full 1–40 Hz range. On noise-free data, NG-RC outperforms spline interpolation across all channels and frequency bands, with the most pronounced improvements observed in the high-frequency β and γ ranges and over frontal and central regions. Under additive Gaussian noise of increasing intensity, the performance of both methods decreases monotonically; however, NG-RC consistently remains more accurate, and the performance gap between the methods widens as noise intensity increases. This advantage is especially evident in the β and γ bands, where spline interpolation deteriorates substantially. These findings suggest that NG-RC provides a robust data-driven alternative for recovering corrupted or missing EEG channels in noise-prone recordings.

Keywords: EEG, ЭЭГ, Reservoir computing, резервуарные вычисления