MRieHy framework improves cross-day BCI decoding with Riemannian hypergraphs

Motor imagery brain-computer interfaces (MI-BCIs) let users control devices by imagining movements, but they often fail across different days because brain signals shift. A new paper introduces Multi-feature Riemannian Hypergraph (MRieHy), a framework designed to adapt online during testing, avoiding the need for recalibration. MRieHy first computes Riemannian means of covariance matrices from cross-day training data to align the distributions across sessions. It then builds a hypergraph over covariance matrices using Riemannian distance, capturing higher-order relationships among samples, and a second hypergraph over deep features using cosine similarity. These are fused with adaptively learned weights, so the model can emphasize the most reliable view at each moment.

During online testing, MRieHy keeps a first-in-first-out buffer of recent samples, applies Riemannian alignment on that buffered data, and decodes using the trained hypergraph. This lets the system continuously adapt to signal drift in real time. The team validated MRieHy on a private four-class ECoG dataset and two public four-class EEG datasets. Results show consistent performance gains over state-of-the-art baselines, with particular improvements in cross-day transferability. The work was led by Siqi Li from Peking University and the Chinese Institute for Brain Research, in collaboration with NeuCyber Neurotech, Beijing Medical University, and Peking Union Medical College. The paper is available on arXiv, and its methods could bring MI-BCI systems closer to practical, daily-use applications like prosthetic control and neural communication.