Visual cues in MI-BCI induce biases in EEG classification models with deep learning but not with standard machine learning.

Objective: Deep learning classification methods are increasingly popular in the brain-computer interface (BCI) field due to their ability to learn complex, non-linear representations from EEG data. Unlike traditional machine learning approaches, which typically rely on hand-crafted features, deep learning models autonomously extract spatio-temporal features directly from the raw or minimally preprocessed signal. However, this flexibility may come at the cost of reduced interpretability and increased susceptibility to confounding factors. This study aims to investigate whether the inclusion of task-related visual cues in the input EEG window biases deep learning models by leading them to exploit irrelevant features, thereby compromising their neurophysiological validity. Approach: This study investigates how including EEG signals measured during visual cue presentation on classification accuracy and feature extraction in motor imagery (MI)-based BCIs. Three classification pipelines were evaluated: CSP+LDA, a traditional machine learning approach; DeepConvNet, a convolutional neural network for EEG classification; and REVE, a Transformer-based EEG foundation model. All methods are trained and tested on two large scale open-access EEG datasets using two epoching strategies, with or without including the visual cue onset, to assess their impact on model performance and neurophysiological relevance. Attribution maps (e.g., Saliency maps) and CSP patterns are used to interpret the features learned by each model. Main results: Including the visual cue onset significantly improved the classification accuracy of DeepConvNet and REVE. However, attribution analysis showed that the model mainly relied on early parietal activations, most likely cue-related visual evoked potentials, rather than MI-related sensorimotor rhythms (ERD/ERS), which were more prominent when the cue onset was excluded. In contrast, CSP+LDA showed minimal accuracy changes across conditions, with consistent motor cortex activation patterns. Significance: These findings demonstrate that higher classification performance does not necessarily imply neurophysiological validity. In particular, deep learning models may exploit confounding signals, such as visual cue processing, which undermines their interpretability and applicability in asynchronous or real-world BCIs, in which there is no such cue. We advocate for systematically applying explainability tools in BCI research to ensure classifiers rely on task-relevant brain activity. Additionally, we emphasize the importance of excluding visual cue periods from EEG segments containing reposenses evoked by the visual cue onset from the input to deep learning models.