Bayesian Neural Field Modeling for Near-Infrared Spectroscopy Driven Brain-Computer Interface Classification of Motor Imagery

Introduction: Brain–Computer Interfaces (BCIs) have emerged as a promising technology for assisting patients with motor impairments. While many BCI systems rely on invasive methods, near-infrared spectroscopy (NIRS) offers a non-invasive alternative for translating brain activity into control signals. Unfortunately, Stroke patients—who would benefit most from this technology—remain an understudied area in noninvasive BCI applications because traditional AI algorithms used in BCIs struggle to accommodate the variability arising from neoplastic changes after a stroke. By leveraging a Bayesian probabilistic framework to manage uncertainty, the present study aims to develop a BCI algorithm that is powerful enough to adapt to the variability of post-stroke neural networks.

Methods: NIRS recordings from 15 stroke patients during BCI motor rehabilitation were processed. Each patient was equipped with NIRS devices featuring 28 channels placed over the brain. These channels captured signals for oxygenated (HbO) and deoxygenated hemoglobin (HbR) as patients imagined performing right and left-hand movements. A Bayesian Spatiotemporal Neural Network (BayesNF) was trained and evaluated to classify motor imagery tasks using a within-subject approach (training = 70%, testing = 30%).

Results: Total BCI exposure per patient averaged 200.53 minutes (SD = 100.77). The BayesNF model achieved an overall mean accuracy of 88% (AUC = 92%), with subject-level accuracies ranging from 72%–98%. In contrast, a support vector machine (SVM) trained on the same dataset yielded a cross-validated accuracy of 46.88%.

Conclusion : The present study achieved an accuracy of 88% (AUC=92%) in classifying motor imagery tasks using NIRS-based BCIs, significantly higher than the standard for non-invasive BCI systems (85-90%). The notable performance in this challenging population—where post-stroke neuroplastic changes often complicate interpretation of cortical signals—underscores the novelty of our approach. By contrast, a standard SVM on the same data yielded markedly lower accuracy (≈47%), reinforcing the value of our BayesNF design over traditional AI models. To our knowledge, this is the first study to achieve a high accuracy in classifying motor movements in stroke survivors, potentially setting a new reference point for future studies.