Research On Deep Learning Methods For The Cross-Dataset Variability Problem In Brain-Computer Interface

Brain-computer interface(BCI)directly converts minds into commands,providing a natural way to interact with the environment.Through decades of effort,BCI has been much more commercialized and is promising to bring the next generation of humanmachine interaction.However,the BCI variability problem hinders practical usages of BCIs,due to electroencephalography(EEG)signals being non-stationary and variable across different subjects.Recently,deep learning has shown advantages in decoding BCI commands,but there is still the generalizability issue of deep learning.This study conducted a series of researches on this problem for BCIs.Firstly,this study conducted many algorithm experiments on multiple motor imagery(MI)and steady-state visual evoked potential(SSVEP)datasets,including within-subject classification,cross-subject classification,and cross-dataset classification experiments.Based on the results of experiments,this study proposed the cross-dataset variability problem in BCIs,in which deep learning models can not generalize well outside the training dataset.The cross-dataset validation scheme is more suitable for evaluating the generalizability of decoding algorithms in actual scenarios.This study also analyzed the main factors that may cause the cross-dataset variability problem for each paradigm.The rest of the study includes algorithm optimization experiments for decoding MI and SSVEP tasks.For MI decoding tasks,this study proposed a general adaptation pipeline for deep learning models,based on reducing the difference between data distributions.The pipeline includes optimization strategies for reducing covariate shift and internal covariate shift problems in BCIs.The results of experiments on 8 MI datasets show that deep learning models with the pipeline significantly outperformed the state-of-the-art benchmark algorithms in the cross-dataset scenario(p < 0.001)and the classification accuracy can be increased by about 7.7%.For SSVEP decoding tasks,this study designed the fixed template network and dynamic template network based on the idea of template matching.The results of experiments on 3 SSVEP datasets show that these two models significantly outperformed the state-of-the-art benchmark algorithms in the within-subject and cross-subject scenarios(p < 0.001)and the average classification accuracy of the BETA dataset can be increased by about 8.8% at 0.5s.Finally,due to the lack of available BCI datasets and algorithms,this study implemented an open-source platform for developing BCI decoding algorithms.This platform includes a dataset processing framework and a decoding algorithm library.The dataset processing framework can realize the data reading and preprocessing pipeline automatically.The decoding algorithm library contains the main BCI decoding algorithms with unified APIs.This platform can process 14 BCI datasets,with 28 decomposition methods,6 Riemannian geometry methods,6 deep learning models,and3 transfer learning methods.The verification results of BCI experiments based on this platform show that this platform can effectively develop BCI decoding algorithms.As a basic research tool,this platform will provide technical support for further researches about the cross-dataset variability problem.In summary,this study verified the generalizability of deep learning models on multiple datasets for BCIs and proposed the cross-dataset variability problem.This study also optimized deep learning models for decoding MI and SSVEP decoding tasks,exploring the possibility of alleviating the BCI variability problems with deep learning models.This study finally built a platform for developing BCI decoding algorithms.The results of this study are promising to improve the generalizability of deep learning models and promote the commercialization of BCI applications in the future.