NAKUL: Spectral-Graph State Space Models for Medical Signal Analysis

The newly introduced model, NAKUL, enhances state space models (SSMs) for the examination of multi-channel physiological signals. It overcomes three primary drawbacks of conventional SSMs: the inability of fixed kernels to represent multi-scale temporal dynamics, Markovian state updates that limit global context, and the neglect of spatial electrode topology through channel-independent processing. NAKUL employs dynamic kernel generation with parallel SSM branches, utilizing varying kernel sizes (3, 5, 7, 11 timesteps) that are weighted by a meta-network for flexible temporal scale selection. Additionally, it features spectral context modeling through FFT-based techniques with learnable Gaussian frequency band filters to identify global periodic patterns at O(N log N) complexity, alongside graph-guided spatial attention based on fixed electrode adjacency. This model is tailored for medical signal analysis, such as EEG or ECG, where temporal dynamics can vary from hundreds of milliseconds in motor preparation to tens of milliseconds during execution transients. The research can be found on arXiv under ID 2605.00871.