cambrain-realtime-eeg-inference

name: cambrain-realtime-eeg-inference description: CaMBRAIN methodology for real-time continuous EEG inference using causal Mamba state space models. First model enabling long-range streaming inference of variable-length EEG signals with >10x higher throughput. version: 1.0.0 author: arXiv paper extraction arxiv_id: 2605.28792 activation_keywords: - EEG - real-time inference - state space model - Mamba - causal model - continuous EEG - streaming inference - brain activity monitoring - neural signal processing tags: - neuroscience - EEG - state-space-models - causal-inference - real-time-processing - deep-learning - computational-neuroscience related_skills: - jet-eeg-flow-matching - eeg-foundation-model-adapters - mamba-spike-forecaster-bci - eeg-ieeg-bridge-bci

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

arXiv ID: 2605.28792 Authors: Abhilash Durgam, Nyle Siddiqui, Jeffrey A. Chan-Santiago, Qiushi Fu, Elakkat D. Gireesh, Mubarak Shah Submission Date: 2026-05-27 Categories: cs.AI, cs.HC, cs.LG

Overview

CaMBRAIN is the first Causal, Mamba-based state space model (SSM) capable of real-time inference of EEG signals. It addresses critical limitations in existing EEG deep learning approaches:

1. Quadratic scaling problem: Traditional attention-based models scale quadratically with sequence length
2. Fixed-length input constraint: Raw EEG must be processed in sliding windows, preventing global signal understanding

Core Innovation

Key Arguments

• Bidirectional approaches are needlessly expensive for EEG processing
• EEG is inherently causal and unidirectional - past signals influence future, not vice versa
• Causal SSM architecture is more appropriate than bidirectional attention

Technical Challenge

EEG exhibits extreme temporal dynamics:

• Crucial events can be extremely brief (fractions of a second)
• Events separated by long intervals (minutes)
• Current self-supervised objectives optimize for signal reconstruction
• These objectives fail to train hidden state to retain salient long-range context

Methodology

Multi-Stage Self-Supervised Training Pipeline

Designed specifically for streaming SSMs to:

1. Encourage long-range memory retention
2. Preserve linear-time complexity of state space models
3. Explicitly train hidden state for streaming inference context

Architecture Components

• Causal Mamba backbone: Unidirectional processing aligned with EEG temporal structure
• Streaming-friendly design: Continuous inference without sliding-window limitations
• Linear-time complexity: O(n) scaling vs O(n²) attention

Performance Results

State-of-the-art across 3 different EEG datasets:

• >10x higher throughput than existing models
• First model enabling long-range, continuous inference of variable-length EEG signals
• Real-time processing capability

Technical Details

Comparison with Existing Approaches

Use Cases

1. Real-time brain activity monitoring
2. Continuous EEG analysis (hours of data)
3. Event detection in long EEG recordings
4. Streaming inference for clinical applications

Implementation Guidance

When to Use

• Long EEG recordings (>seconds to hours)
• Real-time processing requirements
• Memory-constrained environments
• Streaming inference scenarios
• Clinical EEG monitoring

Integration Patterns

1. Replace sliding-window EEG models with streaming inference
2. Combine with EEG foundation models for pre-training
3. Use with event detection pipelines for real-time monitoring
4. Integrate with clinical systems for continuous patient monitoring

Research Context

Related Work

• EEG foundation models (LaBraM, NeuroBERT)
• State space models (Mamba, S4)
• Self-supervised EEG learning
• Brain-computer interfaces

Novel Contributions

1. First causal EEG SSM argument
2. Multi-stage streaming training pipeline
3. >10x throughput improvement
4. Variable-length continuous inference capability

Practical Applications

Clinical

• ICU patient monitoring
• Seizure detection in long recordings
• Sleep stage analysis
• Anesthesia depth monitoring

Research

• Large-scale EEG dataset analysis
• Real-time BCI systems
• Neural dynamics studies
• Brain state tracking

Code & Resources

• Paper: https://arxiv.org/abs/2605.28792
• PDF: https://arxiv.org/pdf/2605.28792v1
• Categories: cs.AI, cs.HC, cs.LG

Key Takeaways

1. Causality matters: EEG is inherently unidirectional - bidirectional models are inefficient
2. Hidden state training: Self-supervised objectives must explicitly train memory retention
3. Linear complexity: Streaming SSMs enable real-time processing at scale
4. Global context: Variable-length inference overcomes sliding-window limitations