Modality-Aware Spectral Alignment for Personalized Federated EEG–Gaze Learning

Objective evaluation of robot-assisted surgery performance can benefit from multimodal physiological and visual measurements, but strong subject heterogeneity limits centralized and standard federated models. In this letter, we propose FedAlign-RAS, a personalized federated framework for multimodal surgical performance prediction using electroencephalography (EEG) and eye-gaze signals. The proposed method combines globally shared EEG and gaze encoders with lightweight client-local spectral and temporal alignment modules to compensate for subject-specific EEG shifts and cross-modal mismatch. We use the PhysioNet EEG-eye-gaze dataset for robot-assisted surgery performance evaluation as the main benchmark and leverage a public consumer-grade EEG–eye-tracking dataset for auxiliary pretraining. Experimental results show that the proposed approach improves performance prediction over centralized, local-only, and standard federated baselines. These findings suggest that subject-specific spectral-temporal personalization is an effective strategy for multimodal surgical skill assessment with raw EEG and gaze data kept local.