J Sleep Med.  2020 Dec;17(2):128-137. 10.13078/jsm.200022.

Diagnostic Accuracy of Different Machine Learning Algorithms for Obstructive Sleep Apnea

• Kim HW
• Park E
• Kim DJ
• Mun SJ
• Kim J
• Lee GH
• Cho JW

• KMID: 2511096
• DOI: http://doi.org/10.13078/jsm.200022

Abstract

Objectives
The objective of this study was to develop models for predicting obstructive sleep apnea (OSA) based on easily obtainable clinical information of patients using various machine learning techniques.
Methods
We used a data set that included the records of 1,368 patients, in which 1,074 patients were male (78.5 %), and 294 patients were female (21.5 %). We randomly divided the data into a training set (1,000) and test set (368). Five machine learning methods, i.e., support vector machine model, lasso logit model, naïve bayes, discriminant analysis, and K-nearest neighbor (KNN), with a 10-cross fold technique were used with the proposed model to predict OSA. We evaluated the accuracy, sensitivity, specificity, and precision of each model for three thresholds [Apnea-Hypopnea Index (AHI)≥5, AHI≥15, and AHI≥30].
Results
Among the machine learning techniques, KNN showed the best results compared to the other techniques. The accuracy, sensitivity, specificity, and precision of OSA prediction were 87.0%, 91.0%, 74.4%, and 91.9%, respectively, based on AHI≥5. When the threshold of OSA was AHI≥15 or AHI≥30, KNN provided lower accuracy (79.6% each) and precision (79.0% and 68.7%), which were still higher than those of the other techniques.
Conclusions
The model derived from the KNN technique exhibited the best performance based on its highest level of accuracy. We demonstrate that this model is a useful tool for predicting OSA.