LuMamba: Latent Unified Mamba for Electrode Topology-Invariant and Efficient EEG Modeling

Abstract

LuMamba (Latent Unified Mamba) is a self-supervised framework combining topology-invariant encodings with linear-complexity state-space modeling: LUNA’s learned-query cross-attention for channel unification and FEMBA’s bidirectional Mamba blocks for efficient temporal modeling. It provides the first systematic investigation of LeJEPA for biosignal learning. Pre-trained on over 21,000 hours of unlabeled EEG, the 4.6M-parameter model attains 80.99% balanced accuracy on TUAB and state-of-the-art Alzheimer’s detection (0.97 AUPR), while requiring 377× fewer FLOPs than state-of-the-art models at equivalent sequence lengths.