Screening for moderate to severe obstructive sleep apnea by using heart rate variability features based on random forest algorithm

Abstract

Purpose More than 80% of patients with moderate to severe obstructive sleep apnea (OSA) are still not diagnosed timely. The prediction model based on random forest (RF) algorithm was established by using heart rate variability (HRV), clinical and demographic features so as to screen for the patients with high risk of moderate and severe obstructive sleep apnea. Methods The sleep monitoring data of 798 patients were randomly divided into training set ( n = 558) and test set ( n = 240) in 7:3 proportion. Grid search was applied to determine the best parameters of the RF model. 10-fold cross validation was utilized to evaluate the predictive performance of the RF model, which was then compared to the performance of the Logistic regression model. Results Among the 798 patients, 638 were males and 160 were females, with the average age of 43.51 years old and the mean body mass index (BMI) of 25.92 kg/m 2 . The sensitivity, specificity, accuracy, F1 score and the area under receiver operating characteristic curve for RF model and Logistic regression model were 94.68% vs. 73.94%; 73.08% vs. 86.54%; 90.00% vs. 76.67%; 0.94 vs. 0.83 and 0.83 vs. 0.80 respectively. Conclusions The RF prediction model can effectively distinguish patients with moderate to severe OSA, which is expected to carry out in a large-scale population in order to screening for high-risk patients, and helps to evaluate the effect of OSA treatment continuously.