Benchmarking EEG Foundation Models for Decoding Alzheimer’s Disease–Relevant Neural Dynamics

Early detection and measurement of Alzheimer’s disease (AD) has the potential to delay disease progression, yet existing fluid-based and imaging diagnostic and monitoring approaches are often costly and typically used only after clinical symptoms appear. Electroencephalography (EEG) offers a low-cost, non-invasive and portable alternative, but substantial inter-subject variability has limited its potential as a practical screening tool. EEG foundation models (EEG-FMs), pre-trained on large-scale unlabeled datasets, offer the prospect of learning representations that generalize across individuals. However, their effectiveness on AD-related tasks remains largely unexplored. In this work, we benchmark six state-of-the-art EEG-FMs against a deep learning model trained from scratch. Using a minimal preprocessing pipeline and subject-independent cross-validation, we evaluate model performance across four levels of difficulty: basic brain states (eyes open/closed), AD detection (classification between AD and control), AD-relevant cognitive task dynamics (executive function, attention and short-term/working memory) and subtle pre-symptomatic genetic risk associated with the ApoE-ε4 allele. We consider both full fine-tuning and frozen linear probing settings. EEG-FMs consistently recover brain states and working-memory load, demonstrating strong transferability. Several models also retain these representations under frozen linear probing. In contrast, prediction of genetic risk remains at chance level across all evaluated models. These findings establish that EEG-FMs can capture AD-relevant functional and pathological signals across unseen individuals and support their further evaluation as scalable biomarkers of AD diagnosis and disease progression in early-stage AD trials.