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

Researchers introduce CaMBRAIN, a causal state space model based on Mamba architecture that enables real-time, continuous EEG signal processing with linear-time complexity. The model achieves state-of-the-art results across multiple datasets while processing signals >10x faster than existing attention-based methods, overcoming critical limitations in handling variable-length brain activity recordings.

CaMBRAIN addresses a fundamental computational bottleneck in EEG analysis by replacing the quadratic-scaling attention mechanisms that dominate current deep learning approaches with a linear-time state space model. Traditional EEG systems require sliding-window processing of fixed-length inputs, fragmenting long recordings and preventing models from capturing critical context across minutes-long intervals. The researchers correctly identify that EEG signals are inherently causal and unidirectional, making bidirectional approaches computationally wasteful.

The technical innovation extends beyond architecture selection. The team developed a multi-stage self-supervised training pipeline specifically engineered to preserve long-range dependencies while maintaining streaming inference capabilities. This addresses the fundamental challenge that brief, salient EEG events—lasting fractions of a second—often occur far apart in time, requiring models to maintain coherent hidden states across extended periods. Prior reconstruction-based self-supervised objectives failed to enforce this explicit long-range memory retention.

The practical implications are substantial for clinical and neuroscience applications. Healthcare providers could monitor continuous EEG streams in real-time without computational constraints, enabling faster seizure detection, sleep stage classification, and neurological event recognition in ICU settings. The >10x throughput improvement makes deployment on edge devices or lower-cost infrastructure feasible. For AI researchers, this work demonstrates that architectural choices matter deeply—state space models prove superior to transformers for this specific domain, challenging the broader assumption that attention mechanisms are universally optimal.

The breakthrough positions streaming neural signal processing as an emerging category worthy of specialized model design rather than generic architecture application. Future work may extend these principles to other physiological signals like ECG or EMG.

• →CaMBRAIN uses causal state space models to process variable-length EEG signals in real-time with linear computational complexity, eliminating quadratic scaling limitations of attention mechanisms.
• →Custom multi-stage self-supervised training enforces long-range memory retention crucial for capturing EEG events separated by extended intervals.
• →Model achieves >10x higher throughput than existing methods while reaching state-of-the-art performance across three independent EEG datasets.
• →Architecture addresses domain-specific constraints by leveraging the causal, unidirectional nature of brain electrical signals rather than applying generic bidirectional approaches.
• →Clinical deployment becomes viable for continuous real-time EEG monitoring in ICUs and neurology applications without computational infrastructure constraints.