Sleep Stage Recognition from EEG Using a Distributed Multi-Channel Decision-Making System

This study investigates the problem of automatic sleep stage recognition using EEG signals collected simultaneously from multiple sensors (electrodes). Given that the multichannel diagnosis does not always lead to an anonymous agreement between diagnostic channels (assessors), the question was how to arbitrate between different assessors to determine a final diagnosis. A multi-channel classification system consisting of interconnected neural networks was tested and compared with the baseline single-channel approaches and a multi-channel method using arbitrary decision-making rules. The system has a distributed problem-solving structure. It included a set of parallel assessor networks performing independent diagnoses based on data collected from different sensors. These individual assessments were assembled and passed to the final decision-making network working as an arbitrator and determining the final diagnostic outcome. The proposed method was validated using publicly available EEG data from the Sleep Disorders Center of the Ospedale Maggiore of Parma, Italy (physionet.org). The presented automatic classification and decision-making approach led to a 98.26% average accuracy, outperforming the baseline single-channel classifiers by 20%-33% and the multi-channel systems with traditional arbitrary decision-making methods by 4%-6%.