Approximate-Inverse Explainability of β–VAE Latents for Multichannel EEG Participant-Generalised Topographical Representation Learning

Electroencephalography (EEG) offers millisecond-level resolution of brain activity. However, its clinical utility is limited by artefacts and its non-linear and non-stationary characteristics. Deep learning has been utilised to model EEG data and learn recurrent, high-level representations, particularly through disentangled representation learning. However, the resulting models and their latent representations are often difficult to interpret because of their unstable and high-dimensional nature. To address this gap, a participant-generalised β weighted Variational Autoencoder (β VAE) is trained on pooled EEG spatially preserved topographic maps from multichannel data. Subsequently, its latent space is explained via HuberAIME, an extension of the model-agnostic Approximate Inverse Model Explanations (AIME) method. In detail, it computes an approximate inverse of the encoder under a Huber loss constraint, providing global feature importance maps for every latent dimension. This is tested against two state-of-the-art explainable AI (XAI) methods, LIME AND SHAP, utilising multichannel EEG data recorded from various participants. The findings demonstrate that HuberAIME produces mutually exclusive and anatomically significant scalp patterns, with each latent axis consistently linked to distinct electrophysiological topographies across all participants. Compared with LIME and SHAP, HuberAIME preserved the distinctiveness of the axis and spatial coherence, ensuring stable and interpretable mappings. In addition, it is significantly faster than LIME which requires several hours and assigns similar importance maps to multiple axes, and it is superior to SHAP, which is computationally prohibitive and produces spatially incoherent explanations. This study contributed to the body of knowledge by proposing an original deep generative learning and interpretability pipeline (β – VAE + HuberAIME), a fast and principled method for linking disentangled EEG latents to concrete, consistent, and organised EEG topographic maps. This enables the reliable and consistent use of learned features across participants, potentially paving new paths for scientific discovery and generalizable diagnostics.