Multi-Channel EEG Analysis for Automated Sleep Staging Characterization Using Ensemble Learning Techniques

Abstract

Advancements in electrical engineering and signal processing can enable the development of automated sleep data analysis systems that embody significantly more advantages than their traditional manual counterparts. Recent approaches to automated sleep stage classification are considered from the PICO framework (Population, Intervention, Comparison, Outcome). Traditionally, PSG/EEG-based sleep studies were a laborious and time-consuming process. Despite this, significant promise is being offered from promising stride development in machine learning (ML) and deep learning (DL) technologies for full automation of sleep stage scoring. The study is applied to data from the SleepEDF public PSG Hypnogram dataset, which combines EOG signal processing and other physiological signals to classify sleep into five stages. The raw signals are segmented into epochs and passed through four machine learning models: Random Forest, Gradient Boosting, Bagging Classifier, and Ensemble Learning. Those accuracies were classified using criteria such as accuracy percentages of 78% for Random Forest, 79% for Gradient Boosting, 75% for Bagging Classifier, and 85% for Ensemble Learning. The comparison has demonstrated Ensemble Learning as the most accurate model and, hence, has shown prospects for being implemented into consumer-grade sleep monitoring devices. Besides highlighting the advantages of automated systems in classification tasks of sleep stages, the results of the experiment suggest possible applications in real-world sleep tracking.