Efficient Calibration in Motor Imagery BCIs Under Data Constraints via Subject Transfer

Tekin Gunasar, Virginia R. de Sa

Keywords: Brain Computer Interface, Riemannian Geometry, EEG

TL;DR: This work explores transfer learning for BCIs with a focus on realistic limited-data settings relevant to online model calibration

Abstract: The practical deployment of motor imagery brain-computer interfaces (MI-BCIs) is bottlenecked by their dependence on large amounts of subject-specific calibration data. To overcome this, we introduce Riemannian Transfer CSP (RTCSP), a transfer learning method that aligns EEG covariance matrices from previous users to a new target subject using Riemannian geometry. This process generates robust spatial filters that perform well even when the target subject's calibration data is severely limited. Our method enables efficient and accurate calibration with minimal data, directly addressing a key challenge for practical MI-BCI systems.

Submission Number: 38