Singapore Institute of Technology

Researchers at the Singapore Institute of Technology decoded online whether 20 healthy subjects were imagining moving their dominant hand left or right, without actually moving, reaching an average accuracy of 58.7±8%, which they say beats existing deep learning models. The difficulty is that neural activity for different directions of the same hand overlaps heavily in space. The model uses EEGNet convolutional filters to extract temporal and spatial features, is trained first as a general model on pooled multi-subject calibration data and then fine-tuned per person, and its squeeze-and-excitation layers rank each electrode's contribution, offering a basis for personalized electrode layouts. Earlier studies mostly compared left versus right hand, largely offline. The work targets stroke rehabilitation but has been tested only in healthy people.