Research map / Time series / Foundation models for time series

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces

NeurIPS 2026 · Time series › Foundation models for time series · approach: Time series foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Benchmarks 10 EEG foundation models, 3 time-series foundation models and supervised EEG baselines across 42 datasets in four domains (epilepsy, sleep, brain age, BCI) using frozen-encoder linear probes and clinically grounded metrics.

Problem

Existing EEG foundation model evaluations use single datasets per domain, single metrics (AUROC/accuracy), and rarely compare against time-series FMs or supervised EEG models, so the real value of EEG-specific pretraining is unclear.

Approach

A unified frozen-encoder linear-probe benchmark spanning many datasets per clinical domain, with model-specific preprocessing matched to each model's pretraining, clinically motivated metrics (event-level sensitivity at fixed false-alarm rates, hypnogram features, brain-age gap, confound-filtered BCI decoding), and three model families compared side by side.

Key results

• Epilepsy: dataset-best Event-Sens@FA AUC ranges from ~0.50 (Helsinki) to ~0.74 (SeizeIT1); NeuroLM and REVE are the only EEG-FMs significantly beating chance top-3 rank, and most models are within ~1 std of the CBraMod random-init baseline (0.44).
• Sleep: Supervised-Pre models lead Cohen's kappa (per-dataset best 0.62-0.85), but the gap largely disappears with sequence length 1, showing sequence modeling drives the advantage.
• Brain age: Sun et al. feature-based model is the only model significant for both MAE improvement and Pearson r; FM-derived brain-age gaps show weak healthy vs. cognitively impaired separation (AUROC ~0.59 at best).

Surprising findings

• Most EEG foundation models perform on par with a randomly initialized CBraMod on seizure detection.
• Motor imagery decoding by FMs is driven largely by non-neural confounds: removing eye/post-cue artifacts drops performance below chance.

Limitations

• Only linear probing evaluated; fine-tuning and few-shot excluded.
• Random-init baseline limited to a single model (CBraMod).
• BrainWave excluded due to incomplete implementation availability; pretraining/downstream dataset overlap prevents clean public scoreboard.

Experimental setup

Datasets: CHB-MIT Siena Helsinki Bonn TUSZ TUAB NMT Epilepsiae SeizeIT1 SeizeIT2 Sleep-EDF Expanded MASS

Benchmarks: NeuroAtlas

Metrics: Event-Sens@FA AUC AUROC balanced accuracy Cohen's kappa MAE Pearson r AUPRC normalized balanced accuracy brain-age gap AUROC

Baselines: CBraMod random-init Seizure-TF DeepSOZ-HEM SleepTransformer SleePyCo CoRe-Sleep Sun et al. brain-age model EEGNet

Builds on: SzCORE CBraMod LaBraM MOMENT

Closest papers by meaning

• What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.91)
• EEG Benchmarking Needs a Task Specification Layer: NeuroDoc for Rulebook-Guided, Executable Benchmark Construction (NeurIPS 2026 · similarity 0.91)
• BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.90)
• NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.90)
• CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.89)
• Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation (NeurIPS 2026 · similarity 0.88)