Unbiased categorical classification of pediatric sleep disordered breathing.

Abstract

To classify pediatric sleep disordered breathing (SDB) using unbiased approaches. In children, decisions regarding severity and treatment of SDB are conducted solely based on empirical observations. Although recognizable entities clearly exist under the SDB spectrum, neither the number of SDB categories nor their specific criteria have been critically defined. Retrospective cohort analysis and random prospective cohort. Community and clinical sample. Urban 5- to 9-year-old community children undergoing overnight sleep study (NPSG), and a comparable prospectively recruited clinical SDB sample. N/a. Principal component analysis was used to identify the uniqueness of the polysomnographically derived measures that are routinely used in clinical settings: apnea-hypopnea index, apnea index, obstructive apnea index, nadir SpO2, spontaneous arousal index and respiratory arousal index. These measures were then incorporated using unbiased data mining approaches to further characterize and discriminate across categorical phenotypes. Of 1,133 subjects, 52.8% were habitual snorers. Six categorical phenotypes clustered without any a priori hypothesis. Secondly, a non-hierarchical model that incorporated 6 NPSG-derived measures enabled unbiased identification of algorithms that predicted these 6 severity-based clusters. Thirdly, a hierarchical model was developed and performed well on all severity-based clusters. Classification and predictive models were subsequently cross-validated statistically as well as clinically, using 2 additional datasets that included 259 subjects. Modeling reached approximately 93% accuracy in cluster assignment. Data-driven analysis of conventional NPSG-derived indices identified 6 distinct clusters ranging from a cluster with normal indices toward clusters with more abnormal indices. Categorical assignment of individual cases to any of such clusters can be accurately predicted using a simple algorithm. These clusters may further enable prospective unbiased characterization of clinical outcomes and of genotype-phenotype interactions across multiple datasets.