Exploring sex-related Biases in Deep Learning Models for Motor Imagery Brain-Computer Interfaces

Abstract

Motor imagery (MI) brain-computer interfaces (BCIs) are promising technologies for neurorehabilitation. In this context, deep learning (DL) models are increasingly being used to decode the mental imagination of movement. However, countless studies across multiple domains have shown that DL models are susceptible to bias, which can lead to disparate performance across subpopulations in terms of protected attributes, such as sex. The reported presence of sex-related information in electroencephalography (EEG) signals, widely used for MI-BCI, further raises warnings in this regard. For this reason, we conducted an in-depth analysis of the performance of DL in terms of the sex and other potential confounding factors. While an initial basic stratified analysis in terms of sex showed differences in favor of the female population, further analysis revealed that performance disparities were actually primarily driven by the discriminability of EEG patterns themselves, and not by the DL model. Moreover, DL models improve overall performance as well as per-group performance, particularly helping subjects with less discriminable EEG patterns. Our work highlights the benefits of DL methods for MI-BCI as well as the need for careful analysis when it comes to bias assessment in complex settings where multiple variables interact. We argue that in-depth studies of model behavior beyond standard performance metrics, should become widespread in the community in order to ensure the development and later deployment of fair BCI systems.