Abstract
Deep learning (DL) models have achieved impressive performance in EEG-based prediction tasks, but they often lack interpretability, limiting their clinical utility. In this study, we introduce a novel self-supervised learning (SSL) framework inspired by neurophysiological reactivity. Our approach models healthy EEG transitions between ocular states by predicting an EEG-derived feature under eyes-open conditions, using features from eyes-closed recordings. The residual between observed and predicted values quantifies devia- tions from normative brain dynamics and serves as a candidate biomarker. To improve clinical relevance, we propose two optimisation strategies that promote residuals predictive of pathology. We evaluated the frame- work using healthy cohorts (LEMON, Dortmund Vital) for self-supervised training, and the AI-Mind cohort for downstream prediction of plasma p-tau217 levels, a proxy for cognitive pathology. Despite extensive hyperparameter optimisation, predictive performance remained poor across all methods, including baseline models, suggesting limitations in the downstream proxy or input signal. Nonetheless, our approach pro- vides a transparent methodology for transition-based EEG biomarker discovery grounded in self-supervised learning. The AI-Mind project aims to develop predictive tools for identifying individuals with mild cognitive impairment (MCI) who are at risk of progressing to dementia.
Authors
Thomas Tveitstøl, Mats Tveter, Christoffer Hatlestad-Hall, Hugo L Hammerc, Ira R J Hebold
https://doi.org/10.1101/2025.09.21.677598