Towards the Correction of Covariate Shift in EEG-Based Passive Brain-Computer Interfaces for Out-of-Lab Applications

The increasing adoption of wearable EEG technology is enabling the development of passive Brain-Computer Interface (pBCI) systems for real-world applications, in the near future, such as Industry 5.0. However, one major challenge in classifying electroencephalographic (EEG) signals in these settings is covariate shift, which occurs when the distribution of the data changes between training and testing sessions due to variations in EEG headset positioning. This study investigates the effectiveness of a linear transformation approach to mitigate the negative effect of covariate shift. Simulations were conducted by using different shift conditions (i.e. deviation of the headset position from the original one), to evaluate (i) the performance of the transformation function used for mitigating the covariate shift occurrence and (ii) the importance that the change of reference and/or channels has on the classification performance. Results show that normalizing covariate shift-affected data (i.e., target) using shift-free data as a template (i.e., source) helps mitigate the negative impact of covariate shift, leading to improved classification performanceThe accuracy loss drops from 14% to 6% in the worst configuration and from 5% to 4% in the best configuration. This improvement is more pronounced when the shift is larger, i.e., when both the reference and channels change between the control dataset and the test dataset. These findings have significant implications for the development of robust and reliable pBCI models for out-of-the-lab contexts.