EEG-DTIL: An EEG-based dynamic task-incremental learning method for decoding ADL-oriented motor imagery pairs

Abstract

Patients with strokes are likely to suffer from dyskinesia and lose their ability to perform activities of daily living (ADL). Motor neurorehabilitation can be gradually realized by continuously learning ADL-oriented motor imagery (MI) pairs with opposite directions, and an electroencephalography (EEG)-based brain-computer interface (BCI) is an effective solution. However, decoding the undetermined streams of MI pairs remains a great challenge for maintaining the balance between new and old tasks. Thus, an EEG-based dynamic task-incremental learning method, called EEG-DTIL, is proposed for decoding progressively incoming MI pairs. Based on the wavelet packet transform, all of the reconstructed subband signals are applied to extract global-view spatial features (GSF) via the regularized common spatial pattern, and the preferred parts are used to capture local-view spatial features (LSF) via tangent space mapping from the Riemannian space. Multiview spatial features (MvSF) are obtained after performing fusion and dimensionality reduction on the GSF and LSF. Inspired by the broad learning system (BLS), a series of personalized sub-BLSs are created in the same order as the inflowing MI pairs and used to construct a residual-based stacked BLS (R-SBLS). Moreover, a dynamic weight consolidation block (DWC) is developed to remember more learned knowledge by controlling the key output weights to be updated within a low error range. Finally, R-SBLS and DWC are combined in parallel, forming a dynamic incremental learning network (DRI-Net). On public and self-collected datasets, EEG-DTIL achieves incremental decoding accuracies of 70.53% and 71.80% for task streams with two and three MI pairs, respectively. The experimental results demonstrate that EEG-DTIL is significantly superior to the related methods, exhibiting better plasticity for new MI pairs, retainability for old MI pairs, and robustness to MI pair sequences.