0459 Diagnostic Performance of Symptomless Obstructive Sleep Apnea Prediction Tools in Clinical and Community-based Samples. (12th April 2019)

• Record Type: Journal Article
• Authors: Holfinger, Steven J; Lyons, Melanie M; Mindel, Jesse W; Cistulli, Peter A; Sutherland, Kate; Chen, Ning-Hung; McArdle, Nigel; Gislason, Thorarinn; Penzel, Thomas; Han, Fang; Li, Qing Y; Mazzotti, Diego R; Keenan, Brendan T; Pack, Allen I; Magalang, Ulysses J
• Is Part Of: Sleep. Volume 42(2019)Supplement 1
• Page Start: A184
• Page End: A185
• Publication Date: 2019-04-12
• 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/zsz067.458
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Most prediction tools for obstructive sleep apnea (OSA) include patient reported symptoms. However, symptoms may not be available in electronic medical records for widespread identification of OSA. Thus, we developed OSA predictions without patient reported symptoms using logistic regression (LOG) or artificial neural network (ANN) and validated their performance in an international clinical sample and in a community-based sample. Methods: Retrospective data on 17, 448 subjects who underwent polysomnography in five international sleep centers within the Sleep Apnea Global Interdisciplinary Consortium were allocated into training (n=10, 469) and validation sets (n=6, 979). Two models to predict the presence of OSA (Apnea Hypopnea Index ≥ 15 events/hour) were developed from the training set based on probabilities derived from LOG and ANN using age, gender, BMI, and ethnicity. Model performance was evaluated using the area under the curve (AUC), positive (PPV) and negative predictive values (NPV), sensitivity and specificity, and positive (+LR) and negative likelihood ratios (-LR). The predictive models were validated in the clinical sample and in the Sleep Heart Heath Study (SHHS) community-based sample (n=5, 761). Results: In the clinical sample validation group, the LOG model had sensitivity=0.65, specificity=0.51, PPV=0.60, NPV=0.56, +LR=1.32, -LR=0.69, and AUC=0.61. The corresponding ANN values were sensitivity=0.74, specificity=0.51, PPV=0.63, NPV=0.64, +LR=1.51, -LR=0.50 and AUC=0.68. The ANN performed significantly better than LOG (p<0.05), except for specificity (p=0.90). In the SHHS validation set, diagnostic characteristics of LOG were sensitivity=0.81, specificity=0.36, PPV=0.25, NPV=0.87, +LR=1.26, -LR=0.53 and AUC=0.63. The corresponding ANN values were sensitivity=0.90, specificity=0.36, PPV=0.28, NPV=0.93, +LR=1.41, -LR=0.27 and AUC=0.72. Again, all ANN characteristics were significantly better than LOG (p<0.05), except for specificity (p=0.74). Conclusion: Compared to LOG, ANN without patient reported symptoms provides improved diagnostic performance for OSA prediction in both clinical and community-based samples. The AUC of the ANN tool was higher than the one previously reported for a tool that includes symptom responses in the same SHHS cohort (STOP-BANG questionnaire). This suggests that the symptomless ANN prediction may also have utility in identifying OSA risk but future studies in other samples are warranted. Support (If Any)