Development and validation of moderate to severe obstructive sleep apnea screening test (ColTon) in a pediatric population.

ABSTRACT

OBJECTIVE: Development and validation of a machine learning algorithm to predict moderate to severe obstructive sleep apnea syndrome (OSAS) in otherwise healthy children. DESIGN: Multivariable logistic regression and cforest algorithm of a large cross-sectional data set of children with sleep-disordered breathing. SETTING: An university pediatric sleep centre. PARTICIPANTS: Children underwent clinical examination, acoustic rhinometry and pharyngometry, and surveying through parental sleep questionnaires, allowing the recording of 14 predictors that have been associated with OSAS. The dataset was nonrandomly split into a training (development) versus test (external validation) set (21 ratio) based on the time of the polysomnography. We followed the TRIPOD checklist. RESULTS: We included 336 children in the analysis: 220 in the training set (median age [25th-75th percentile] 10.6 years [7.4; 13.5], z-score of BMI 1.96 [0.73; 2.50], 89 girls) and 116 in the test set (10.3 years [7.8; 13.0], z-score of BMI 1.89 [0.61; 2.46], 51 girls). The prevalence of moderate to severe OSAS was 106/336 (32%). A machine learning algorithm using the cforest with pharyngeal collapsibility (pharyngeal volume reduction from sitting to supine position measured by pharyngometry) and tonsillar hypertrophy (Brodsky scale), constituting the ColTon index, as predictors yielded an area under the curve of 0.89, 95% confidence interval [0.85-0.93]. The ColTon index had an accuracy of 76%, sensitivity of 63%, specificity of 81%, negative predictive value of 84%, and positive predictive value of 59% on the validation set. CONCLUSION: A cforest classifier allows valid predictions for moderate to severe OSAS in mostly obese, otherwise healthy children.