Progressive transfer learning Unifies Multi-Corpus EEG for robust and scalable Imagined-Speech decoding

Imagined-speech brain-computer interfaces (BCIs) face limitations due to limited, subject-specific data and the risk of catastrophic forgetting during sequential adaptation. We present a gated progressive neural network (PNN) that assimilates information from diverse imagined-speech EEG datasets and applies it to new users. The signals are processed through a standard pipeline that includes 1–70 Hz band-pass filtering, baseline correction, per-channel z-normalisation, and ICA, followed by a comprehensive, physiologically motivated feature vector that encompasses spatial (CSP), time–frequency (discrete wavelet statistics), spectral (δ–γ band powers), and time-domain dynamics (Hjorth parameters, autocorrelation, higher-order moments). The PNN evolves by incorporating a new column for each task; previously established columns remain unchanged while lateral adapter layers and a shared gate control how prior learned embeddings influence the new column, effectively preventing forgetting while allowing for reuse. Initially, the model trains on ASU, Kara One, and FEIS, then adapts to 15 new subjects from BCI Competition 2020 IV (BCIC2020-IV). The pooled-corpus accuracies are 99.83 % (ASU), 95.77 % (Kara One), and 99.73 % (FEIS). For BCIC2020-IV, the progressive transfer achieves a median accuracy of 93.7 % (ranging from 76 % to 95 %), compared to 54.0 % (ranging from 42 % to 72 %) for subject-specific baselines, with correspondingly improved macro-F1 scores and a more organised, class-balanced structure in the learned embeddings. Ablation studies and permutation tests reveal that the autocorrelation and wavelet blocks play key roles. These results demonstrate that progressive transfer learning lessens the calibration requirements for each subject while reducing the chances of forgetting, offering an effective and scalable approach to developing robust imagined-speech BCIs.