Sleep Stage Classification: Investigating Deep Learning with High Dataset Imbalance Under Temporal Interval Change and Data Imbalance Techniques

This study investigates automated sleep stage classification using a CNN-GRU architecture, demonstrating that three-channel multimodal data and longer temporal contexts (60s and 90s) improve classification performance compared with two-channel and 30s approaches. The findings support ASSC as an efficient and accurate alternative to manual sleep scoring, with the best-performing model achieving a macro F1 score of 0.78 ± 0.02.

Sleep stage scoring is the process of assigning sleep stages to periods of sleep depending on features of electrophysiological signals. Automated Sleep Stage Classification (ASSC) using machine learning aims to improve efficiency of manual sleep scoring by experts, since sleep stage scoring is tedious and time-consuming when done manually. This study aims to investigate deep learning techniques through CNN-GRU architecture, primarily aiming to investigate the effects of temporal interval of data and statistical techniques to address dataset heavy imbalance. Three channels (two EEG plus EOG) was found to be superior to two channels (only two EEG) in all models tested. A 60s temporal interval data two-stage bypass model with N1 data augmentation utilizing 3 channels was found to have the best average performance with macro f1 score 0.78 ± 0.02 and macro N1 f1 score 0.51 ± 0.03, but certain 90s temporal interval models and two-stage variants showed statistically similar performance. The results reaffirm the superiority of a multimodal approach (three channel vs two channel) and the importance of context (60s and 90s vs 30s). It is evident that sleep stage scoring can be conducted efficiently and accurately through ASSC.

Explore More!

Research Paper