Riemannian Transfer Learning in Motor Imagery decoding: Reproducibility and Standardized Benchmarks

Motor imagery (MI)-based brain-computer interfaces (BCIs) hold significant potential for rehabilitation and assistive technologies. However, their widespread adoption is hindered by high inter-subject variability in electroencephalogram (EEG) signals, necessitating extensive calibration for new users. Transfer learning (TL) methods overcome this by leveraging data from existing subjects to reduce the calibration time. However, the lack of standard evaluation protocols in EEG-MI TL research makes it challenging to compare different approaches fairly. Moreover, the lack of availability of codebases adds to the issue of reproducibility. In this paper, we propose a standardized evaluation protocol to compare key transfer learning techniques across cross-session and cross-subject scenarios. We further conduct ablation studies focusing on signal length and preprocessing parameters to quantify the sensitivity of the algorithms to signal and noise variability. Finally, we present Python implementations of the methods for reproducibility.