0495 Prediction Of Obstructive Sleep Apnea Using Machine Learning Technique. (27th April 2018)

• Record Type: Journal Article
• Authors: Huang, W; Lee, P; Liu, Y; Lai, F
• Is Part Of: Sleep. Volume 41(2018)Supplement 1
• Page Start: A186
• Page End: A186
• Publication Date: 2018-04-27
• Subjects: Sleep -- Physiological aspects -- Periodicals; Sleep disorders -- Periodicals; Sommeil -- Aspect physiologique -- Périodiques; Sommeil, Troubles du -- Périodiques; Sleep disorders; Sleep -- Physiological aspects; Sleep -- physiological aspects; Sleep Wake Disorders; Psychophysiology; Electronic journals; Periodicals; 616.8498
• DOI: 10.1093/sleep/zsy061.494
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Many questionnaires or prediction models tried to identify patients with obstructive sleep apnea (OSA) to prioritize sleep study. However, the performance was not unified which was related to the prevalence and definition of OSA, characteristics of participants, and feature selected. Therefore, the present study proposed a novel model to predict OSA with the minimal features. Methods: We collected the clinical features and polysomnographic parameters from 5, 301 patients (mean age 47.5 ± 14.4 y/o, men 76.5%) referred for suspect OSA [apnea-hypopnea index (AHI) ≥ 10/hr]. Among 5, 301 patients, the mean AHI was 29.7 ± 26.1/hr and 70.1% had OSA. The patient numbers of training set and testing set for building model were 3, 048 and 2, 253, respectively. Thirty-five features including comorbidities, anthropometric information, and symptom were utilized to investigate the relevance to OSA. AHI≥10/hr were considered to suffer from OSA. Support Vector Machine (SVM) was employed for the judgement of feature effectiveness during selection procedure. To eliminate class-imbalance effect, training samples were re-sampling by Synthetic Minority Over-sampling Technique. Eventually, a model to predict OSA was designed based on SVM and the selective set of features. The performance of OSA prediction models were assessed with area under receiver operating characteristic curve (AUROC), sensitivity, specificity, accuracy, Likelihood ratio-positive [LR (+)], and Likelihood ratio-negative [LR (-)]. Results: With 35 features, the AUROC, sensitivity, specificity, accuracy, LR (+), and LR (-) was 0.818, 85.6%, 58.2%, 77.2%, 2.05 and 0.25, respectively. Nine features had been selected based on model performance non-inferior to that in 35-features AUROC. The 9-feature model gave AUROC, sensitivity, specificity, accuracy LR (+), and LR (-) as 0.818, 86.5%, 57.4%, 77.5%, 2.03 and 0.24, respectively. Conclusion: The 9-feature model had a good performance for identifying patients with OSA. It could be applied to prioritize patients for polysomnography which may bring earlier treatment. Support (If Any): The study was supported by grants from National Taiwan University (number NTU-ERP-104R8951-1, 105R8951-1, 106R880301, 106R880302) and NTU-NTUH-MediaTek Innovative Medical Electronics Research Center.