Evaluating the biological meaning of neural network decisions in EEG-based MCI detection.

Abstract

High accuracy in medical classification tasks does not ensure that neural networks reason in ways consistent with clinical or neurobiological understanding. This study examines whether a Vision Transformer (ViT) trained on resting-state EEG infers cognitive impairment through physiologically meaningful mechanisms. A lightweight ViT was trained on multi-center resting-state EEG to detect mild cognitive impairment. The model's probabilistic outputs were interpreted as continuous cognitive risk scores. Knowledge distillation and spatial perturbation analyses were performed to identify the electrophysiological features and cortical regions underlying the model's predictions. The model achieved an average accuracy of 75.4% in five-fold cross-validation, and generalized to Alzheimer's disease cohorts and an external clinical center. The derived risk scores correlated with MoCA subdomains, particularly memory, language and orientation. Key drivers included increased autocorrelation, reduced Lempel-Ziv complexity and changes in power spectral density. Perturbation analyses highlighted strong contributions from the insular cortex and the transverse temporal regions. The model's decision process reflects physiologically and anatomically interpretable patterns consistent with clinical reasoning, supporting EEG-based modeling as an objective tool for quantifying cognitive function.