CaMBRAIN: Real-time, Continuous EEG Inference with Causal State Space Models

Problem
This paper addresses the limitations of existing deep learning models for EEG signal processing, particularly their reliance on attention mechanisms that exhibit quadratic scaling with sequence length. This scaling issue hampers the ability to process long-duration EEG signals effectively. Additionally, current methods typically require fixed-length inputs, which necessitates a sliding-window approach that compromises the model’s ability to capture global context. The authors propose CaMBRAIN, a novel Causal Mamba-based state space model (SSM) designed for real-time, continuous EEG inference. This work is presented as a preprint and has not yet undergone peer review.

Method
CaMBRAIN employs a causal state space model architecture that leverages a multi-stage self-supervised training pipeline. This pipeline is specifically designed to enhance long-range memory retention, which is critical for capturing brief yet significant EEG events that may be separated by long intervals. The model maintains linear-time complexity, a significant advantage over traditional attention-based models. The training process focuses on optimizing the hidden state to retain salient long-range context, which is essential for streaming inference. The authors do not disclose specific details regarding the training compute or the exact datasets used for training and evaluation.

Results
CaMBRAIN achieves state-of-the-art performance across three distinct EEG datasets, demonstrating over 10x higher throughput compared to existing models. While specific numerical results are not provided in the abstract, the authors claim that their model significantly outperforms named baselines, although the exact metrics and comparisons are not detailed. The emphasis on throughput suggests a focus on real-time applicability, which is a critical factor for practical EEG applications.

Limitations
The authors acknowledge that training a model capable of capturing brief EEG events is inherently challenging due to the sporadic nature of these events. They also note that while their approach improves long-range context retention, it may still struggle with extremely rapid fluctuations in EEG signals. Additionally, the paper does not address potential issues related to the generalizability of the model across different populations or EEG recording conditions, which could impact its applicability in diverse clinical settings.

Why it matters
The development of CaMBRAIN has significant implications for real-time EEG analysis, particularly in clinical and research settings where continuous monitoring is essential. By enabling long-range, continuous inference of variable-length EEG signals, this model could facilitate advancements in neurofeedback, seizure detection, and brain-computer interfaces. The ability to process EEG data in real-time with high throughput opens new avenues for applications that require immediate feedback and analysis, potentially transforming how EEG data is utilized in both clinical and experimental contexts.

Authors: Abhilash Durgam, Nyle Siddiqui, Jeffrey A. Chan-Santiago, Qiushi Fu, Elakkat D. Gireesh, Mubarak Shah
Source: arXiv:2605.28792
URL: https://arxiv.org/abs/2605.28792v1