A decision support system for automatic sleep staging from HRV using wavelet packet decomposition and energy features

Abstract

Heart rate variability (HRV) can reflect the relationship between heart rhythm and sleep structure. In order to achieve long-term effective monitoring of sleep and to solve the generalization problem of sleep staging algorithms, HRV signals were used to identify the wake, non-rapid eye movement (NREM) and rapid eye movement (REM) stages of two databases with 11,597 epoches of 16 subjects (8 healthy subjects and 8 sleep disorder). Features were extracted from HRV using three different methods: Hilbert Huang Transform (HHT), Singular Value Decomposition (SVD), and Wavelet Packet Decomposition (WPD). The extracted features were reconstructed into a new sleep score table, classified using random forest (RF) classifier and the results were interpreted using basic statistical criteria. Among these methods, the WPD extraction feature is the most successful. For the sleep staging model trained in the first database, the accuracy of Wake, NREM and REM after WPD extraction features were 80.9%, 88.2% and 65.8% respectively. The second databases were tested using a trained model, in which the Avg.Acc, Avg.Precision, Avg.F and Kappa statistic for healthy subjects were 72.9%, 0.720, 0.716 and 0.439 ± 0.0346. The standard explanations for sleep disorder subjects were 77.8%, 0.780, 0.843 and 0.656 ± 0.0324. The experimental results show that the energy characteristics of WPD extraction are the best. In addition, there was a significant difference in Kappa values (p < 0.05) between healthy subjects and sleep disorder subjects by variance analysis of standard interpretation, and it can be used as a basis for judging whether a subject has a sleep disorder.