Template-Anchored Signal Decomposition and Label Refinement for Impairment Categorization on Noisy Visual Electrophysiological Data.

Visual electrophysiology offers an objective alternative to subjective vision assessments, providing enhanced reliability but remaining underutilized for visual impairment categorization due to imperfect datasets, noisy signals, and label inconsistencies. To address these challenges, we first construct a new dataset comprising ERG and VEP recordings from 1,970 participants with various ocular conditions and a mean age of 21.59 $\pm$ 4.47 years. A three-class categorization task is defined to reflect clinical disability grading. We then propose a template-anchored decomposition method for noise-aware signal feature extraction, using mode-specific ridge regression to isolate pathological components by removing both healthy patterns and noise artifacts. To collectively enhance feature representation from multi-channel signals, we design a joint attention encoder that combines channel attention with bilateral spatial attention across left- and right-eye channels to emphasize informative and coherent signal patterns. Furthermore, we introduce a template-anchored refinement mechanism that addresses label noise by updating soft labels during training based on the distances between each feature vector and the class templates. Extensive experiments demonstrate that our method achieves consistent improvements over existing electrophysiology-based and general biosignal classification methods, supporting its feasibility for visual impairment categorization in a cohort composed primarily of young adults.