Visual impairment categorization using dictionary-decomposed electrophysiological data on a dual-path state-space convolution framework

This is the authors' abstract. We don't add key points for this paper.

Accurate visual impairment categorization is crucial in forensic and social welfare applications, while conventional grading based on visual acuity and perimetry is subjective and susceptible to unreliable patient cooperation. Visual electrophysiology offers an objective, tamper-resistant alternative, but its application to precise impairment grading is hindered by limited datasets, coupled pathological patterns, and insufficient modeling ability. To address these issues, we construct a complete multimode visual electrophysiology dataset comprising two electroretinography (ERG) modes and one visual evoked potential (VEP) mode, including 1369 subjects with verified impairment gradings. We further propose a dictionary decomposition method that separates normative structural deviations from raw signals by projecting them onto a convolutional dictionary learned from healthy templates while preserving residual components that may be related to uncharacterized disease-specific anomaly patterns. To exploit both global and multiscale local temporal characteristics, a dual-path state-space convolution framework is designed, where the state-space branch captures global trends and the convolution branch extracts multiscale local features via learnable wavelet encoding. A two-stage cross-attention finally facilitates heterogeneous feature fusion. Extensive experiments on the proposed dataset demonstrate clear performance gains over existing visual electrophysiology-based and general biosignal classification methods, with the proposed framework achieving a median balanced accuracy (BACC) of 76.52%, a recall of 74.90%, and an F1 score (harmonic mean of precision and recall) of 73.42%, highlighting its effectiveness for objective and reliable visual impairment categorization.