BrainWave: A Brain Signal Foundation Model for Clinical Applications (2402.10251v6)

Abstract: Neural electrical activity is fundamental to brain function, underlying a range of cognitive and behavioral processes, including movement, perception, decision-making, and consciousness. Abnormal patterns of neural signaling often indicate the presence of underlying brain diseases. The variability among individuals, the diverse array of clinical symptoms from various brain disorders, and the limited availability of diagnostic classifications, have posed significant barriers to formulating reliable model of neural signals for diverse application contexts. Here, we present BrainWave, the first foundation model for both invasive and non-invasive neural recordings, pretrained on more than 40,000 hours of electrical brain recordings (13.79 TB of data) from approximately 16,000 individuals. Our analysis show that BrainWave outperforms all other competing models and consistently achieves state-of-the-art performance in the diagnosis and identification of neurological disorders. We also demonstrate robust capabilities of BrainWave in enabling zero-shot transfer learning across varying recording conditions and brain diseases, as well as few-shot classification without fine-tuning, suggesting that BrainWave learns highly generalizable representations of neural signals. We hence believe that open-sourcing BrainWave will facilitate a wide range of clinical applications in medicine, paving the way for AI-driven approaches to investigate brain disorders and advance neuroscience research.

• The paper introduces Brant-2, a foundation model featuring 1 billion parameters and trained on 4TB of mixed EEG and SEEG data to advance brain signal analysis.
• It employs temporal and spatial encoders within a multi-FFN Transformer block and uses dual pretraining tasks like mask prediction and forecasting to enhance performance.
• Empirical evaluations demonstrate robust results in seizure detection, sleep stage classification, emotion recognition, and motor imagery, even in low-resource clinical scenarios.

Advancing the Frontier of Brain Signal Analysis with Brant-2: A Comprehensive Foundation Model

Introduction

The burgeoning interest in brain signal analysis is motivated by its potential to unlock novel insights into brain function, disease mechanisms, and innovative applications in neurology, sleep studies, and even human-computer interaction. The landscape of brain signal research, straddling both invasive (e.g., SEEG) and non-invasive (e.g., EEG) techniques, presents unique challenges, notably the massive diversity in data and the prohibitive costs of data annotation. This context sets the stage for the significant contribution of foundation models—models pretrained on large, unlabeled datasets that can be fine-tuned for specific tasks with minimal labeled data. In this discourse, we explore Brant-2, the latest in foundation models for brain signals, marking a significant leap in addressing the multifaceted challenges of this domain.

Model Overview

Brant-2 emerges as a progressive iteration over its predecessor, Brant, with substantial advancements in its pre-training corpus, robustness to data variations, and adaptability across a diverse range of tasks. Featuring a staggering 1 billion parameters and trained on nearly 4 TB of mixed SEEG and EEG data from over 15k subjects, Brant-2 embodies a strategic fusion of volume and diversity in its training regime.

Architectural Innovations

Key to Brant-2's architecture is the incorporation of temporal and spatial encoders within a multi-FFN (Feedforward Neural Network) Transformer block, enabling the model to master both the spatial correlations between electrodes and the temporal dependencies across signals. This layered architectural strategy is complemented by a sophisticated data augmentation module, enhancing the model's resilience to data variations and modeling scales. Additionally, two-pretraining tasks—mask-prediction and forecasting—equip Brant-2 with the ability to extract rich semantic information, setting a solid foundation for its generalization capabilities.

Empirical Evaluation

Extensive evaluations underscore Brant-2’s superior performance across a spectrum of downstream tasks, including but not limited to seizure detection and prediction, sleep stage classification, emotion recognition, and motor imagery classification. Brant-2 consistently outperforms prior models and achieves competitive results against task-specific methods, demonstrating its broad applicability and robust generalization prowess.

Scalability and Efficiency in Low-Resource Scenarios

Our scalability analysis reveals a proportional improvement in model performance with an increase in parameters, affirming the potential of Brant-2 to scale adeptly. Furthermore, in scenarios with scarce labels, Brant-2's performance exhibits remarkable stability in comparison to fully supervised methods. This characteristic is particularly beneficial in clinical settings, where obtaining extensive labeled data is often impractical, thus highlighting the model's practical utility and the significant reduction in dependency on labeled data.

Conclusion and Future Directions

Brant-2 stands as a testament to the power of foundation models in the field of brain signal analysis, setting a new benchmark for both academic research and practical applications. Looking ahead, the ambition is to further enrich Brant-2’s pre-training corpus, incorporating a wider array of data sources to extend its applicability across more research areas and real-world scenarios. The promising results of Brant-2 not only fortify the foundation model paradigm in brain signal analysis but also pave the way for new avenues of exploration and innovation in understanding the enigmatic workings of the human brain.