Simplified feature extraction for low-resource sleep staging

Wearable sleep monitoring devices require efficient algorithms capable of running on resource-constrained hardware. This study develops a lightweight approach to sleep stage classification optimized for low-power environments. We first analyze the computational complexity of standard EEG feature extraction methods, then design simplified approximations that maintain discriminative power while significantly reducing computational requirements. Our progressive computation framework calculates basic features first, only proceeding to more complex features when classification confidence is low. Experiments on the Physionet sleep EEG dataset demonstrate that our approach achieves 93.2% of the accuracy of full-complexity methods while reducing power consumption by 76% and memory usage by 68%. Model compression techniques, including 8-bit quantization and network pruning, further optimize performance on microcontroller-class hardware. The system successfully classifies sleep stages with only 32KB of RAM and 120KB of flash memory, enabling integration into wearable devices with minimal battery impact. This lightweight methodology makes continuous, long-term sleep monitoring feasible in real-world settings without sacrificing clinical utility.