Generative Diffusion-Based Task Incremental Learning Method for Decoding Motor Imagery EEG.

BACKGROUND/

objectivesMotor neurorehabilitation can be realized by gradually learning diverse motor imagery (MI) tasks. EEG-based brain-computer interfaces (BCIs) provide an effective solution. Nevertheless, existing MI decoding methods cannot balance plasticity for unseen tasks and stability for old tasks. This paper proposes a generative diffusion-based task Incremental Learning (IL) method called GD-TIL.

methodsFirst, data augmentation is employed to increase data diversity by segmenting and recombining EEG signals. Second, to capture temporal-spatial features (TSFs) from different temporal resolutions, a multi-scale temporal-spatial feature extractor (MTSFE) is developed via integrating multiscale temporal-spatial convolutions, a dual-branch pooling operation, multiple multi-head self-attention mechanisms, and a dynamic convolutional encoder. The proposed self-supervised task generalization (SSTG) mechanism introduces a regularization constraint to guide MTSFE and unified classifier updating, which combines labels and semantic similarity between the augmentation with original views to enhance model generalizability for unseen tasks. In the IL phase, a prototype-guided generative replay module (PGGR) is used to generate old tasks' TSFs by training a lightweight diffusion model based on the prototype and label of each task. Furthermore, the generated TSF is merged with a new TSF to fine-tune the convolutional encoder and update the classifier and PGGR. Finally, GD-TIL is evaluated on a self-collected ADL-MI dataset with two MI pairs and a public dataset with four MI tasks.

resultsThe continuous decoding accuracy reaches 80.20% and 81.32%, respectively. The experimental results exhibit the excellent plasticity and stability of GD-TIL, even beating the state-of-the-art IL methods.

conclusionsOur work illustrates the potential of MI-based BCI and generative AI for continuous neurorehabilitation.