A Novel Sleep Stage Contextual Refinement Algorithm Leveraging Conditional Random Fields

Bufang Yang,Wenxuan Wu,Yitian Liu,Hongxing Liu

DOI: https://doi.org/10.1109/tim.2022.3154838

IF: 5.6

2022-01-01

IEEE Transactions on Instrumentation and Measurement

Abstract:Automatic sleep stage classification has gained much attention in recent researches. Various classification algorithms have been proposed for automatic sleep staging, including deep neural networks and traditional machine learning models. However, the output of those models has unreasonable sleep stage transitions, as temporal dependence of sleep stage label of adjacent data segment is ignored. In this article, we propose a novel sleep stage contextual refinement algorithm based on conditional random fields (CRFs). The algorithm works as a post-processing step to rectify the hypnogram produced by sleep staging pre-classifiers. Unreasonable sleep stage transitions can be corrected via our algorithm to further improve the classification performance. We use CNN-based, CNN-LSTM-based, random forest, and two existing sleep staging models UTSN and UTSN-L as pre-classifiers. Our algorithm is evaluated on three sleep datasets, Sleep-EDF-20, DRM-SUB, and SVUH-UCD datasets. Results demonstrate that our CRF contextual refinement algorithm can improve the classification performance of five sleep staging pre-classifiers, including overlapping-based and nonoverlapping-based models, and the algorithm works both on healthy subjects and patients with sleep disorder. When using CNN as pre-classifiers, our algorithm improves the overall accuracy and macro F1-score by 2.5% and 4.7% on Sleep-EDF-20, by 3.6% and 6.6% on DRM-SUB, and by 5.5% and 7.8% on the SVUH-UCD dataset.