POC-CSP: a novel parameterised and orthogonally-constrained neural network layer for learning common spatial patterns (CSP) in EEG signals

<i>Objective</i>. Common spatial patterns (CSPs) has been established as a powerful feature extraction method in EEG signal processing with machine learning, but it has shortcomings including sensitivity to noise and rigidity in the value of the weights. Our goal was to transform CSP into a trainable machine learning model that can learn from data, be regularized, and be integrated into end-to-end classification networks.<i>Approach</i>. We developed a novel parameterised and orthogonally-constrained neural network layer for learning CSPs (POC-CSP) that maintains CSP's mathematical properties while allowing trainable weights. The layer uses parameterisation based on Lie Group theory to convert constrained optimisation into unconstrained optimisation, enabling integration with standard neural network (NN) training methods. We evaluated the approach on two public motor imagery datasets, focusing on both subject-specific and multi-subject paradigms.<i>Main results</i>. POC-CSP outperformed both conventional CSP and existing NN implementations in subject-specific classification tasks. In a novel multi-subject paradigm, POC-CSP achieved superior generalisation. When fine-tuned with just 50% of a new subject's data, POC-CSP achieved 0.95 average accuracy across subjects, substantially outperforming subject-specific models trained with more data.<i>Significance</i>. These findings demonstrate that combining CSP's proven effectiveness with NNs' flexibility can significantly improve EEG signal processing performance. The ability to generalize across subjects and achieve high accuracy with minimal subject-specific training data makes POC-CSP particularly valuable for practical brain-computer interface applications, where collecting large amounts of training data from each new user is often impractical or unfeasible.