S-132 | Exploring Demographic Biases in Deep Learning Models for Motor Imagery BCI

Theoretical and Computational Neuroscience
Author: Bruno Jose Zorzet | Email: bzorzet@sinc.unl.edu.ar

Bruno J. Zorzet^1°, Victoria Peterson^2°, Diego H. Milone^1°,  Rodrigo Echeveste^1°

^1° Research Institute for Signals, Systems and Computational Intelligence, sinc(i), (FICH-UNL / CONICET), Santa Fe, Argentina
^2° Institute of Applied Mathematics of the Litoral, IMAL, (FIQ-UNL / CONICET), Santa Fe, Argentina

Brain-computer interfaces (BCI) are systems that allow direct communication between the brain and external devices. Brain activity signals, like electroencephalography, are captured and processed to be converted into controlling commands. Recently, the use of deep learning methods for decoding the mental states has gained attention. Besides accurate decoding performance, it is essential to ensure that the model behaves fairly across different demographic groups in the data. Here, we study whether the performance of models in motor imagery (MI) tasks are influenced by protected attributes of the users. To assess the presence of biases in decoding MI based on biological sex of the participants, we conducted a rigorous analysis of performance of a deep learning model in an across-subject MI-BCI scenario. Our analysis reveals that certain subjects achieve consistently high performance independent of the training sets and initialization of the models. Furthermore, upon comparing metrics between males and females, we observed a tendency where females outperform males, with significant differences in one of the used datasets. This serves as a warning to the community about the potential presence of biases in across-subject MI-BCI.