BCIwiki (bciwiki.com) — A deep learning framework decoded whether healthy subjects were imagining left or right movements of the same hand, reaching an average online accuracy of 58.7±8% across 20 people, researchers at the Singapore Institute of Technology reported in Sensors (Basel, Switzerland) on October 1, 2026 (PMID 42817507, DOI 10.3390/s26185951). The authors say the result outperforms existing deep learning models.
The work is aimed at stroke rehabilitation. Hemispheric strokes impair motor control on the opposite side of the body, and motor imagery-based brain-computer interfaces (BCIs) are thought to promote neuroplasticity. According to the paper, most deep learning studies decode bilateral imagery, such as left hand versus right hand, and evaluate offline. Decoding direction within a single hand would offer a more natural control interface with more degrees of freedom, but the underlying neural activity overlaps spatially.
The model uses EEGNet-based convolutional filters to extract temporal and spatial features, with squeeze-and-excitation (SE) layers that rank electrode importance and feature maps. A subject-independent model was trained on calibration data from multiple subjects, then fine-tuned for each individual and evaluated on that person's online session data.
The authors say the SE-layer ranking reveals each electrode's contribution and could support subject-specific BCI optimization. All subjects were healthy; the 58.7% average accuracy should be read against a 50% chance level for this binary task, and performance in stroke patients has not been tested.