Toward Reliable passive BCI in Out-Of-The-Labs Environments

The transition of passive Brain-Computer Interfaces (pBCIs) from controlled laboratory settings to real-world applications requires robust systems with built-in self-quality assessment. Covariate shift, often induced by headset displacement, is a critical factor of system failure in real-world applications. This study proposes a two-stage framework for EEG signal management in ecological contexts. First, an Isolation Forest model for unsupervised monitoring, effectively detects anomalies between source and target sessions that correlate with signal quality degradation (r=−0.69). Second, a corrective function based on Riemannian geometry. Experimental results show that this transformation enhances classification performance, achieving AUC values for mental workload detection, that are not significantly different from the source dataset (p<0.05). These findings highlight the framework's potential for maintaining pBCI reliability in non-stationary industrial environments.