Research Models & Releases·arXiv cs.LG·14h ago

EEG foundation models lag classical baselines in clinical dementia tasks

A systematic evaluation of six pretrained EEG foundation models reveals significant gaps between their clinical promise and real-world performance. When tested on dementia classification across multiple datasets, models like REVE substantially underperform classical signal-processing baselines, with frozen transfer accuracy dropping to 0.568 AUROC versus 0.769 for traditional features. The work exposes how dataset identity leakage and population shift undermine generalization claims, using rigorous negative controls including label permutation and random initialization. This benchmarking effort matters because it challenges the narrative that foundation models automatically transfer to clinical domains, forcing practitioners to reconsider whether pretrained representations actually capture clinically relevant EEG patterns or merely memorize dataset artifacts.

Modelwire context

Skeptical read

The paper doesn't just benchmark six models; it systematically dismantles the assumption that freezing pretrained weights transfers clinical utility. The real finding is methodological: dataset identity leakage (where models exploit spurious correlations tied to which dataset they're tested on) explains much of the performance gap, not fundamental limits of the architectures themselves.

This echoes the constraint-focused framing from the EchoBridge and Earth observation papers published the same day. Like EchoBridge's work on long-tail cardiac pathologies, this EEG study exposes how standard evaluation hides failure modes that matter clinically (dementia detection across populations). Both papers reject the premise that a single model trained on aggregate data generalizes; both demand explicit testing for real-world brittleness. The geospatial best-practices paper from July 27 makes a similar point about architectural maturity lagging behind model performance claims.

If the authors release code and the same frozen-transfer AUROC gap (0.568 vs 0.769) reproduces when independent teams retrain REVE and BENDR on their own dementia cohorts, the leakage hypothesis holds. If the gap shrinks below 0.65 AUROC with domain-adaptive fine-tuning on just 50 subjects, that signals the models do capture transferable patterns but the frozen-weight assumption was the error, not the architectures.

Coverage we drew on

EchoBridge: Long-Tail-Aware ECG-Echocardiography Text Alignment for Echocardiography-Derived Cardiac Findings · arXiv cs.LG

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsLaBraM · EEGMamba · CBraMod · REVE · BENDR · BIOT

Read full story at arXiv cs.LG →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Stress-Testing EEG Foundation Models for Clinical Decoding: Dataset Identity and Targeted Negative Controls”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.