Feature selection for sleep staging using cardiorespiratory and movement signals

Abstract

EEG based sleep staging is commonly conducted at clinical setting, which may disturb patients’ sleep habits and thus impair study results. A non-invasive method of sleep staging through cardiorespiratory signals and body movement allow us to classify the stages awake, light, deep and REM sleep using random forest (RF) with good clinical accuracy. The aim is to improve the latter by tuning the RF hyperparameters. Statistical features of size p=63 extracted from vital signals from 13 nights of healthy subjects were used as inputs to the classifiers and classified using 30s epochs. The hyperparameters were tuned over the splitting criteria Gini and entropy, maximal tree depth (up to fully grown), number of trees (up to 1000) and maximal number of features considered at each split (p, vp or log p). Classification accuracies when employing a 10-fold cross-validation were highest with the hyperparameters Gini, vp used features, tree depth of 30 and 1000 trees, yielding an accuracy of (72.8±1.3)%. The feature importance ranking was consistent between the different classifiers, where respiration variability standard deviation always came first with (5.3±2.3)%, ahead of the second by (1.9±1.1)%. Selecting only the most important features may allow to increase the accuracy further by reducing noisy inputs while decreasing computation time. Cardiorespiratory features came out as much more relevant than movement, which indicates that the latter may be omitted without risking a meaningful decrease in scoring accuracy.