NeuroSleep: neuromorphic event-driven single-channel EEG sleep staging for edge-efficient sensing

Area of Science:

• Biomedical Engineering
• Neurological Monitoring
• Edge Computing

Background:

• Continuous neural sensing is crucial for long-term health monitoring on wearable edge platforms.
• High-frequency electroencephalography (EEG) processing for sleep monitoring is computationally intensive and energy-demanding.

Purpose of the Study:

• To propose NeuroSleep, an integrated system for energy-efficient sleep staging on wearable edge devices.
• To address the computational bottleneck in EEG-based sleep monitoring.

Main Methods:

• NeuroSleep utilizes Residual Adaptive Multi-Scale Delta Modulation (R-AMSDM) for event-driven EEG sensing, creating complementary multi-scale bipolar event streams.
• A hierarchical inference architecture with Event-based Adaptive Multi-scale Response (EAMR), Local Temporal-Attention Module (LTAM), and Epoch-Leaky Integrate-and-Fire (ELIF) modules is employed.

Main Results:

• NeuroSleep achieved 74.2% accuracy with 0.932M parameters on single-channel EEG data.
• The system reduced sparsity-adjusted effective operations by 53.6% compared to dense processing.
• NeuroSleep improved accuracy by 7.5% while reducing computational load by 45.8% versus a dense Transformer baseline.

Conclusions:

• NeuroSleep offers a deployment-oriented framework for single-channel sleep staging.
• The system couples neuromorphic event encoding with state-aware context modeling.
• It reduces redundant high-rate processing and enhances energy scalability for wearable and edge platforms.