Machine Learning-Based Decoding of Executed and Imagined Finger Movements from OPM-MEG Data

Sammanfattning

Introduction: Brain-Computer Interfaces (BCIs) have many applications, including enabling communication and supporting motor rehabilitation. Decoding fine movements, such as finger movements, allows for more flexible BCI control. While various neuroimaging modalities have been employed in BCI systems, invasive techniques carry substantial risks, while commonly used non-invasive techniques like electroencephalography (EEG) are often limited by low signal quality. Optically pumped magnetometer magnetoencephalography (OPM-MEG) is a recently developed non-invasive modality that offers several advantages over traditional magnetoencephalography (MEG), including increased data quality, improved adaptability, and reduced costs. However, due to its novelty, OPM-MEG is currently underexplored in BCI research. Research Question: Therefore, this study investigates OPM-MEG’s potential for BCI applications by examining to what extent machine learning models can use OPM-MEG data to decode finger movements. Method: A novel OPM-MEG dataset was collected from a participant performing motor and motor imagery finger movements using four fingers. Four machine learning models, SVM, EEG-SimpleConv, EEGNet, and DetachROCKET Ensemble, were trained and evaluated on this dataset. Results: The models achieved classification accuracies ranging from 65.2–74.9% on the motor dataset and 52.2–60.7% on the motor imagery dataset, all substantially exceeding the chance level of 25%. Discussion: Despite this study’s limited data, the results demonstrate that OPM-MEG enables reliable decoding of finger movements. Performance falls within the range of prior studies, exceeding some results, matching others, and falling below the best-performing approaches, with further improvements expected with the addition of more data. Overall, this study highlights OPM-MEG’s potential as a powerful neuroimaging modality in the context of BCI research and applications.