Collaborative optimization of time periods and frequency bands based on group feature selection for MI-BCI illiteracy

Objective: To improve the classification accuracy of the issue that suboptimal signal segments hinder effective brain activity representation of motor imagery-based brain–computer interface (MI-BCI) illiteracy. Methods: A collaborative group feature selection framework is proposed to jointly optimize cross-domain time periods and frequency bands. Segmented periods are combined with frequency bands to create a “time period-frequency band” analysis matrix. The collected EEG is preprocessed based on this, followed by feature extraction to obtain the feature matrix. Feature selection is used to scan features with good performance so as to assess the signal quality of each time period-frequency band. Specially, to overcome the limitation of previous single-element feature selection methods, a group feature selection technique is used based on fixed units. Building on this foundation, supervised Fisher scoring is adopted in the training domain to identify time period-frequency band combinations with higher scores. Then, an unsupervised group feature selection approach in the testing domain refines the selected combinations for both domains. Results: Using the “BMI-open dataset,” the average classification accuracy for all participants was 72.8%, outperforming the best-performing control method by 1.9%. In the BCI illiteracy group, the average classification recognition accuracy reached 62.0%, registering a 2.9% improvement over the best-performing control method. Conclusion: The proposed method significantly enhanced classification accuracy, particularly for BCI illiteracy users. Significance: This approach systematically addresses key challenges, including the effects of low-quality signal segments and the inconsistencies between optimal time–frequency segments across domains, caused by the non-stationarity and poor repeatability of EEG.