0482 PERFORMANCE OF AN INTERNATIONAL SYMPTOMLESS PREDICTION TOOL FOR OBSTRUCTIVE SLEEP APNEA USING ARTIFICIAL NEURAL NETWORK

• Published in: Sleep (New York, N.Y.) Vol. 40; no. suppl_1; p. A180
• Main Authors: Mindel, JW, Maislin, G, Keenan, B, McArdle, N, Singh, B, Chen, N, Pack, F, Mazzotti, DR, Bittencourt, L, Sutherland, K, Cistulli, P, Pack, AI, Magalang, U
• Format: Journal Article
• Language: English
• Published: US Oxford University Press 28.04.2017
• Subjects: Neural networks; Sleep apnea
• ISSN: 0161-8105; 1550-9109
• DOI: 10.1093/sleepj/zsx050.481

Abstract Introduction: Current prediction tools for obstructive sleep apnea (OSA) include responses to questions about patient symptoms within subjects from a single country. We developed and determined the diagnostic performance of a symptomless OSA prediction tool in a large number of subjects seen in the member centers of the Sleep Apnea Global Interdisciplinary Consortium (SAGIC). Methods: 12,073 patients aged ≥18 years and referred for diagnostic in-laboratory polysomnography (PSG) for suspicion of OSA were included in the study from the following SAGIC centers: Perth, Australia (n=3,904); Columbus, OH (n=5,852); Philadelphia, PA (n=1,053); Taoyuan, Taiwan (n=1,264). A generalized regression artificial neural network (ANN) was used to generate the prediction tool (SAGICNet), which produced the desired output of OSA presence or absence. Variables chosen as inputs were: age, gender, body mass index (BMI), neck collar size, and self-reported ethnicity. The ANN was trained in 8,451 (70%) subjects randomly selected from the dataset and validated in the remaining 3,622 (30%). Results: Subjects (55% male) were 48.8 ± 14.0 years-old with an apnea hypopnea index (AHI) of 27.7 ± 29.6 events/hour; the overall OSA prevalence (AHI≥15/hr) was 54%. The diagnostic characteristics of the symptomless SAGICNet for predicting the presence of OSA in the validation group were: sensitivity (Sens) = 0.75, specificity (Spec) = 0.58, positive predictive value (PPV) = 0.67, negative predictive value (NPV) = 0.67, +Likelihood ratio (+LR) = 1.78, -Likelihood ratio (-LR) = 0.43, and area under the receiver-operator-curve (AUC) = 0.734. For predicting the presence of severe OSA (AHI>30/hr), values in the validation group were: Sens = 0.35, Spec = 0.90, PPV = 0.64, NPV = 0.73, +LR = 3.49, -LR = 0.73, and AUC = 0.723. Conclusion: The symptomless SAGICNet has a +LR comparable to previously reported tools using patient reported symptoms in predicting the presence of moderate OSA. It has high specificity (but low sensitivity) for predicting the presence of severe OSA. The symptomless SAGICNet may be a useful tool for identification of OSA risk in electronic medical records or databases for clinical and research purposes in the international setting. Support (If Any):