Hierarchical Multimodal Sleep Staging with Optimized EEG, EOG, and PPG Features for Wearable Applications
TL;DR
A lightweight two-stage Deep Learning framework for five-class sleep staging based on optimized multimodal physiological features extracted from electroencephalogram (EEG), electrooculogram (EOG), and photoplethysmography (PPG) signals, suggesting its potential suitability for wearable and edge-based sleep-monitoring systems.
Abstract
Automatic sleep staging is fundamental for diagnosing sleep disorders and enabling long-term sleep monitoring with wearable devices. Although Deep Learning has significantly improved classification performance, balancing accuracy with computational efficiency remains challenging, particularly for resource-constrained systems. This paper proposes a lightweight two-stage Deep Learning framework for five-class sleep staging based on optimized multimodal physiological features extracted from electroencephalogram (EEG), electrooculogram (EOG), and photoplethysmography (PPG) signals. The framework is trained and tested using the Bitbrain Open Access Sleep (BOAS) database, considering a 31-subject dataset partitioned into training (24 subjects) and independent test (7 subjects) sets. Feature selection is performed using the minimum Redundancy Maximum Relevance (mRMR) algorithm, followed by Principal Component Analysis (PCA) for EEG and EOG features, while respiratory and cardiac features derived from PPG are directly incorporated into the multimodal representation. A hierarchical Long Short-Term Memory (LSTM) architecture first classifies sleep into Wake, REM, and NREM, then further distinguishes the N1, N2, and N3 stages. On an independent test set, the classifier achieves 88.2% five-class accuracy on the multimodal feature set (EEG + EOG + PPG) with a model size of 3.14 MB, and 87.1% accuracy on the EEG-only feature set using only 2.95 MB of memory. Leave-One-Subject-Out (LOSO) cross-validation yields 86.9% accuracy, supporting subject-independent generalization. Inference latency ranged from 2.55 ms (EEG-only) to 4.02 ms (multimodal), with measured energy per inference of 2.88–10.9 mJ across feature sets. Additional validation on the RichSleep and ISRUC datasets demonstrates robustness across different recording conditions, achieving mean accuracies of 78.2% and 77.3%, respectively. The proposed framework provides a favorable trade-off among classification performance, complexity, and memory footprint, suggesting its potential suitability for wearable and edge-based sleep-monitoring systems.