Leveraging Generic Time Series Foundation Models for EEG Classification

AI summary

Overview

Research area: Machine learning for biomedical signals — specifically, applying generic time series foundation models to electroencephalography (EEG) classification.

Technical level: Intermediate. The paper assumes familiarity with transfer learning, fine-tuning, contrastive pretraining, and standard EEG benchmarks (motor imagery, sleep staging), but the core argument is accessible without deep domain expertise.

Scope: The paper evaluates whether a general-purpose time series classification foundation model (Mantis), pretrained on non-EEG or purely synthetic data, can match or beat an EEG-specific foundation model (CBraMod) and a classic EEG convolutional baseline (EEGNet) on motor imagery and sleep stage classification.

What This Paper Is About

EEG datasets are typically small, fragmented, hard to share, and highly variable across subjects and sessions, which limits how well deep learning models generalize. Meanwhile, foundation models pretrained on large, diverse corpora have transformed vision and language, and similar models are now emerging for time series — but their usefulness for neural signals is largely unexplored.

This paper asks a direct question: is EEG-specific pretraining actually necessary, or can a generalist time series foundation model transfer effectively to EEG? The authors take Mantis, a time series classification foundation model, and test it both with pretraining on heterogeneous real-world time series data and with pretraining on purely synthetic data, then compare against EEGNet and CBraMod on two motor imagery datasets and eight sleep staging datasets.

Key Contributions

1. First systematic evaluation of a generic time series classification foundation model on EEG. The authors apply Mantis (Feofanov et al., 2025) to motor imagery classification and sleep stage prediction, two tasks the model was not designed for.

2. Comparison of two pretraining regimes for EEG transfer. They test Mantis checkpoints pretrained on a mix of real-world time series datasets (1.8 million samples, with only a small portion of EEG data) and on a purely synthetic dataset generated by the CauKer algorithm (1 million samples), alongside random initialization.

3. Head-to-head benchmarking against EEG-specific baselines. Mantis is compared with EEGNet (a tailored CNN with fewer than 0.01M parameters) and CBraMod (the most recent EEG-specific foundation model, 4M parameters), across 2 BCI datasets and 8 sleep staging datasets, with results averaged over 3 seeds.

4. Evidence that domain-specific pretraining may not be required, plus a report of practical fine-tuning requirements. The authors report that freezing the encoder causes a large performance drop, making fine-tuning necessary, and that MOMENT was excluded because its large model size compared to CBraMod and Mantis makes it very difficult to fine-tune.

Main Findings

• Mantis matches or beats EEG-specific models on motor imagery. On PhysioNet-MI, CBraMod reaches 64.17 ± 0.90 balanced accuracy while Mantis with real-data pretraining reaches a bolded 64.43 ± 1.50, and CBraMod keeps the best Cohen's Kappa (52.22 ± 1.70 versus Mantis's 52.13 ± 1.87). EEGNet scores 58.14 ± 1.25 balanced accuracy and 44.68 ± 1.20 Cohen's Kappa.

• Synthetic-only pretraining wins on SHU-MI. Mantis pretrained purely on CauKer-generated synthetic data achieves a bolded 72.15 ± 3.8 AUC-PR versus CBraMod's 69.88 ± 0.07, and a bolded 65.5 ± 4.3 balanced accuracy versus CBraMod's 63.70 ± 1.50 and EEGNet's 58.89 ± 1.77. The authors note this is notable because Mantis was pretrained with minimal EEG data.

• Channel-independent design is an advantage for low-channel tasks. Mantis processes channels univariately and models their inter-dependencies only at the final classification layer, whereas CBraMod relies on multivariate pretraining. Mantis nonetheless outperforms the multivariate approach on the multivariate BCI tasks.

• Both foundation models beat EEGNet on sleep staging, and Mantis consistently beats CBraMod. On the 2-channel sleep setup, gains of Mantis over CBraMod range from 0.3% on CCSHS to nearly 3% on the MASS dataset. Per-dataset weighted F1 (averaged over 3 seeds) is reported as

Authors’ abstract

Foundation models for time series are emerging as powerful general-purpose backbones, yet their potential for domain-specific biomedical signals such as electroencephalography (EEG) remains rather unexplored. In this work, we investigate the applicability a recently proposed time series classification foundation model, to a different EEG tasks such as motor imagery classification and sleep stage prediction. We test two pretraining regimes: (a) pretraining on heterogeneous real-world time series from multiple domains, and (b) pretraining on purely synthetic data. We find that both variants yield strong performance, consistently outperforming EEGNet, a widely used convolutional baseline, and CBraMod, the most recent EEG-specific foundation model. These results suggest that generalist time series foundation models, even when pretrained on data of non-neural origin or on synthetic signals, can transfer effectively to EEG. Our findings highlight the promise of leveraging cross-domain pretrained models for brain signal analysis, suggesting that EEG may benefit from advances in the broader time series literature.