Towards reliable out-of-lab pBCIs: riemannian geometry framework for mitigating covariate shift in real-world passive brain-computer interfaces

Passive brain-computer interfaces (pBCIs) are promising tools for real-world, out-of-lab monitoring of cognitive states (e.g. mental workload). However, their reliability (e.g. classification accuracy) is often compromised by covariate shift caused by changes in headset positioning among different sessions. This study introduces a Riemannian-based correction method de signed to mitigate the negative effect of the shift. We evaluated the framework by training a random forest classifier on a source domain to classify two mental workload levels and then tested on a target domain (rep resenting a channels shift) both before and after applying the transformation. The validation involved three levels of analysis: (i) classification performance via area under curve (AUC) analysis, (ii) statistical assessment of man ifold alignment through Riemannian distance, and (iii) correlation between geometric correction and AUC re covery. The application of the correction significantly improved workload classification, significantly over coming AUC scores compared to the uncorrected target data. Statistical analysis revealed a significant reduction in Riemannian distance between the source and target manifolds after correction (𝒑 = 𝟐.𝟑𝟔𝒆−𝟓). Furthermore, correlation analysis showed a strong positive relationship between geometric alignment and performance gain (Pearson correlation 𝑹 = 𝟎.𝟔𝟐𝟖)