Recursive sample weighted – N-way partial least squares for brain-computer interface decoder online learning with class imbalance

Real-time online decoder learning is effective for Brain-Computer Interfaces (BCIs), enabling the consideration of sensitive responses in the brain neural signals. However, class imbalance can occur during decoder learning. Some tasks / classes being harder to learn than others. If unaddressed, this class imbalance can significantly decrease the decoder’s overall performance. This paper introduces a novel algorithm called Recursive Sample Weighted – N-way Partial Least Squares (RSW-NPLS) to tackle this issue. In the RSW-NPLS, weights are assigned to each learning sample. To do this, the algorithm dynamically tracks the size of class and computes weights to balance minority classes with the majority class at each update. A constraint is applied to avoid large weights during the decoder learning. The RSW-NPLS algorithm is well suited for BCI systems. It addresses several key constraints of BCI: it supports multiclass classification, is designed for online learning, has low computational requirements, and can handle high-dimensional data. The proposed RSW-NPLS algorithm is compared to the Recursive Exponentially Weighted – N-way Partial Least Squares (REW-NPLS), another algorithm adapted for BCI systems but which does not address class imbalance. Both algorithms are evaluated in offline and pseudo-online configurations using three different databases of ECoG signals from paraplegia and tetraplegia patients implanted with WIMAGINE sensors. The RSW-NPLS outperforms the REW-NPLS in term of classification performance and stability. Tests demonstrate that the RSW-NPLS effectively compensates for class imbalance in real-time and maintains strong performance even with significant imbalances.