A Novel Approach to Prediction of Mild Obstructive Sleep Disordered Breathing in a Population-Based Sample: The Sleep Heart Health Study

Abstract

1641Prediction of Mild Obstructive SDB-Caffo et al IN THE ROUTINE CLINICAL EVALUATION OF PATIENTS FOR POSSIBLE SLEEP DISORDERED BREATHING, PHY-SICIANS MAKE JUDGMENTS AS TO THE LIKELIHOODthat a particular person is affected and decide whether polysomnography is warranted based on this likelihood. Their judgments draw on elements of the patient's history, such as reports of snoring or excessive daytime somnolence, demographic factors, and weight or body mass index (BMI). If the estimated likelihood is sufficiently high, then the physician may consider ordering a polysomnogram (PSG). Similarly, media outlets and medical societies put forward recommendations to the public for self-referral to a primary care or sleep physician for the evaluation of possible sleep disordered breathing. For example, in response to the recent publication of the authors' manuscript on sleep and mortality, 1 CBS news recommended that patients discuss with their physicians the need for polysomnography in the presence of severe snoring, restless sleep, frequent waking up at night, and excessive daytime exhaustion.2 Faced with continued patient concern about having sleep disordered breathing, clinicians need tools to identify those individuals most likely to have sleep disordered breathing so as to avoid unneeded referral for a PSG.At the core of this discussion are two questions related to the prediction of the eventual result of a PSG in the general population: first, how predictable is the result of a PSG, and second, what easily collected variables are most influential for predicting the result of a PSG?A number of previous studies have considered whether the respiratory disturbance index (RDI) or the apnea-hypopnea index (AHI) can be predicted with sufficient certainty using such clinical information; these findings reviewed below generally indicate limited predictive power.In this manuscript, we consider the prediction of RDI in participants in the Sleep Heart Health Study (SHHS) 3 using modern ensemble learning algorithms to make predictions based on classifications of predictor variables. 4 Ensemble learning algorithms base their predictions on a collection (ensemble) of simpler prediction equations. We focus on the prediction of mildly elevated RDI values, because patients with severe disease are presumably easier to diagnose without the aid of screening prediction algorithms. The goals of this manuscript are 3-fold: The first is to investigate the cross-sectional prediction of RDI in a population-based sample using these modern ensemble learning algorithms, which sort the data based on patterns of clustering of classifiers.4 A related second goal is to investigate the overall predictability of RDI in this population. A third goal is to demonstrate the utility of ensemble learning algorithms to sleep researchers, as these methods are potentially useful for reducing large amounts of data, such as polysomnographic data, to their most informative elements.This study differs from other previous analyses that focused on individual predic...