Single-Subject Unsupervised Detection of Covariate Shift in EEG: A Step Towards Real-World passive Brain-Computer Interface

Thanks to significant advancements in wearable EEG technology, the application of passive Brain-Computer Interface (pBCI) systems is now becoming feasible in settings beyond the laboratory, such as Industry 5.0, where human machine interaction is central. However, in these real-world and less controlled settings, a common challenge in classifying EEG signals is the variation in EEG headset positioning across sessions, which can induce a phenomenon known as covariate shift, leading to reduced accuracy between training and subsequent recording sessions. This study aims to detect covariate shift, using an unsupervised Machine Learning model (i.e. Isolation Forest), at the single subject level for online applications, focusing on evaluating how quickly the model identifies anomalous records that deviate from the distribution of the control dataset. 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 detection model used for identifying the occurrence of covariate shift and (ii) the number of needed data samples to provide a statistically significant detection. The results show that it is possible to detect Covariate Shift in real-time, using less than 20 seconds of recording. This result was obtained in 9 out of 10 participants, across all test configurations and it indicates that with only 20 seconds of recording, the number of anomalies detected by the method is statistically significant compared to the number of anomalies in a control dataset, with a significance level of 5%. Only in one participant the method fails to reach this level of accuracy for two out of three test configurations. This pilot study represents an important step towards the generation of robust and reliable pBCI models for applications in real-world contexts.