Interpretability of EEG Models via Sparse Autoencoders

Deep neural decoders have demonstrated strong performance in electroencephalography (EEG)-based motor-imagery (MI) brain–computer interfaces (BCIs), yet their internal representations remain difficult to interpret, limiting neuroscientific insight and model auditability. This paper presents a mechanistic interpretability framework that uses sparse autoencoders (SAEs) to decompose the latent representations of MI-EEG decoders, instantiated on CIACNet, a representative decoder and investigates whether individual sparse features encode physiologically meaningful and causally effective correlates of left- versus right-hand motor imagery. Subject competence is first established using the training session (Session-T) data, after which a single shared encoder is jointly trained on the pooled Session-T data from the retained cohort to obtain a common cross-subject latent representation. A label-free Top-K SAE is subsequently trained to decompose this representation into a sparse feature dictionary. Candidate features are interpreted through convergent evidence. This evidence spans discriminative selectivity, cross-subject consistency, sensorimotor μ/β physiology and contralateral topography, causal interventions, semantic specificity, and held-out cross-session feature-level replication on the evaluation session (Session-E). By integrating these complementary lines of evidence, the proposed interpretability framework connects internal decoder representations with their discriminative role, causal influence, and underlying sensorimotor physiology. Features supported across this evidence chain are identified as physiologically consistent and functionally relevant MI representations. The resulting spectral, spatial, and causal characterisation provides a systematic pathway from latent feature discovery to mechanistic understanding of deep EEG models. These findings support the feasibility of the proposed Top-K SAE framework for mechanistic analysis of MI-EEG representations, without implying superiority over alternative decomposition methods or SAE configurations. Our findings highlight the potential of SAEs for developing EEG-BCI decoders that are not only predictive, but also interpretable, physiologically consistent, and auditable across subjects and sessions.