Are foundation models useful feature extractors for electroencephalography analysis?

TL;DR

The paper examines whether general time-series foundation models can serve as effective EEG feature extractors in data-limited clinical contexts. By evaluating OTiS across age prediction, seizure detection, and EEG-event classification, and comparing it to EEG-specialised baselines under zero-shot, linear probing, and fine-tuning regimes, the study shows that OTiS can extract high-quality EEG features and sometimes achieve state-of-the-art performance. It also demonstrates that biomarker localisation is feasible via frequency-band analysis and that model architecture, particularly context length, materially affects diagnostic accuracy. The results suggest that foundation-model–based EEG analysis can reduce reliance on large domain-specific datasets, with practical implications for clinical deployment, while highlighting the importance of targeted preprocessing and adaptation for task-specific success.

Abstract

The success of foundation models in natural language processing and computer vision has motivated similar approaches for general time series analysis. While these models are effective for a variety of tasks, their applicability in medical domains with limited data remains largely unexplored. To address this, we investigate the effectiveness of foundation models in medical time series analysis involving electroencephalography (EEG). Through extensive experiments on tasks such as age prediction, seizure detection, and the classification of clinically relevant EEG events, we compare their diagnostic accuracy with that of specialised EEG models. Our analysis shows that foundation models extract meaningful EEG features, outperform specialised models even without domain adaptation, and localise task-specific biomarkers. Moreover, we demonstrate that diagnostic accuracy is substantially influenced by architectural choices such as context length. Overall, our study reveals that foundation models with general time series understanding eliminate the dependency on large domain-specific datasets, making them valuable tools for clinical practice.

Figures (7)

• Figure 1: Overview. We study the potential of general foundation models to extract demographic and disease-related information from electroencephalography signals.
• Figure 2: First two principal components of zero-shot EEG features extracted by OTiS. The model captures distinct features across frequency bands, enabling the localisation of demographic and disease-specific biomarkers in clinical practice.
• Figure 3: LEMON results. (Left) OTiS$\bullet$ outperforms OTiS$_\text{EEG}$$\bullet$ and other specialised models $\bullet$. (Right) Optimal domain adaptation of OTiS for EEG analysis is achieved through fine-tuning. EEG broadband signals contain the most age-related information, with the information density increasing at higher frequencies.
• Figure 4: Epilepsy results. (Left) OTiS$\bullet$ outperforms OTiS$_\text{EEG}$$\bullet$ and is competitive with other specialised models $\bullet$. (Right) OTiS extracts meaningful EEG features even without domain adaptation. EEG broadband signals contain the most ictal-related information, with no clear frequency-based trend in information density.
• Figure 5: TUEV results. (Left) OTiS$\bullet$ outperforms OTiS$_\text{EEG}$$\bullet$ and other specialised models $\bullet$, except for huge models like LaBraM. (Right) While OTiS is competitive with specialised models in the zero-shot setting, optimal domain adaptation is achieved through fine-tuning. EEG broadband signals contain the most clinically relevant information, with the information density increasing at lower frequencies.