The Illusion of Accuracy in Rehabilitation BCI: Beyond CNN Artifacts

Abstract

This research addresses a critical flaw in current BCI models: the tendency of Convolutional Neural Networks (CNNs) to achieve high accuracy by learning non-neural artifacts (like ocular or muscular noise). Using Saliency Maps and Representational Similarity Analysis (RSA), we demonstrate that Transformers focus on actual neural features, providing a safer and more reliable path for clinical neuro-rehabilitation.

Key Findings:

• Identification of “False Accuracy” in traditional BCI models.
• Superiority of Transformer architectures in multi-class imagined speech decoding.
• Implementation of XAI techniques to validate neural feature selection.