Polysomnography and Treatment-Related Outcomes of Childhood Sleep Apnea

The rationale, design, trial procedures, and results related to the CHAT study have been described by Marcus et al⁸  and Redline et al.⁵  The salient aspects of the trial methodology are provided in the following sections.

The CHAT study was designed to test the hypothesis that adenotonsillectomy is superior to watchful waiting in improving outcomes related to childhood OSA. The study investigators enrolled children on the basis of age (5–9 years) and polysomnographic definitions (AHI >2 or obstructive apnea index [OAI] >1 per hour) based on suitability for adenotonsillectomy. The rationale for the use of these definitions, based on normative data representing the burden of upper airway obstruction, are provided in the trial design.⁵  Children were excluded if they had polysomnographic evidence of very severe OSA (AHI >30¹⁰  or OAI >20) or prolonged hypoxemia. The trial data were obtained from the National Sleep Research Resource (https://sleepdata.org) through a data use agreement.¹¹  The primary and secondary outcomes were measured at baseline and follow-up at 7 months. Table 1 summarizes these outcomes, along with their respective domains and the interpretation of the results.

The primary outcome was the attention and executive function score on the Developmental Neuropsychological Assessment (NEPSY).¹²  Behavioral outcomes were measured by the Conners’ Rating Scales–Revised: Long Version Global Index,¹³  the Behavior Rating Inventory of Executive Function (BRIEF),¹⁴  the Conners’ Comprehensive Behavior Rating Scale (attention-deficit/hyperactivity disorder),¹³  and the Child Behavior Checklist (CBCL).¹⁵  Both global and disease-specific quality-of-life measures were obtained by using the Pediatric Quality of Life Inventory (PedsQL)¹⁶  and the Obstructive Sleep Apnea-18 (OSA-18) scale.¹⁷  The symptoms of OSA were assessed by the Pediatric Sleep Questionnaire Sleep-Related Breathing Disorder (PSQ-SRBD)¹⁸  scale and the modified Epworth Sleepiness Scale (ESS).¹⁹  Generalized intellectual functioning was determined by using the Differential Ability Scales-II.²⁰ 

Additional assessments included changes in BMI percentile,²¹  systolic and diastolic blood pressure measurements (measured in millimeters of mercury), and serum levels of C-reactive protein (measured in micrograms per milliliter) with a detection threshold of 0.15 μg/mL, a nonspecific marker for inflammation.²²  Changes in the Homeostasis Model Assessment for Insulin Resistance (HOMA-IR), a surrogate marker of insulin resistance, was also calculated by using the equation HOMA-IR = fasting insulin (measured in microunits per milliliter) × fasting glucose (measured in milligrams per deciliter)/405.²³  The rationale for the choice of these variables is based on the disease domains potentially affected by untreated OSA.⁵  In the current study, polysomnographic resolution of OSA was defined by an AHI <2 and OAI <1 at follow-up, identical to the CHAT study.^24–26  The change in polysomnographic severity of OSA was the difference between the follow-up and baseline AHI scores. These mediators were chosen on the basis of (1) the finding that half of the children in the nonsurgical arm of the trial experienced resolution of the condition supporting the conclusion that surgery could be potentially avoided in these children,⁸  and (2) change in AHI indicates change in OSA severity, which was prespecified in the trial design as a mediator for other outcomes.⁵  Additionally, a more robust definition of OSA resolution represented by the resolution of obstruction (follow-up AHI <1.5²⁷  and OAI <1) along with the absence of hypoxemia (no portion of sleep with oximetry reading <92%) was also evaluated as a mediator for other outcomes.

The characteristics of the study population including age, sex, race, ethnicity, BMI percentile score, socioeconomic characteristics, and OSA severity defined by the AHI score were compared between 2 groups of children defined by whether they experienced resolution of OSA at follow-up. We then compared the 18 outcomes between the trial arms by performing an analysis of covariance adjusted for the stratification factors of age, race, and weight status.⁸ 

We performed a mediation analysis by modeling the effect of treatment on each of the 18 outcomes in Table 1 by including each of the mediating variables separately. The mediator represents the pathway through which the intervention influences the change in outcome. Causal mediation analysis was described by Imai et al²⁸  on the basis of the well-known conceptual framework of Baron and Kenny.²⁹  Methodologic aspects of the analysis are described in the Supplemental Information. The analysis was undertaken by using the mediation package for R (https://cran.r-project.org, version 3.5.1).³⁰  The first mediation model described in the current study decomposes the total effect of treatment into mediated effect via polysomnographic resolution of OSA and direct effect unrelated to the mediator, as shown in Fig 1. The second model similarly uses the change in AHI as a mediator.

A key assumption supporting the framework of causal mediation analysis is sequential ignorability. In the first part of the assumption, the mediator is randomly assigned given the treatment as part of a randomized trial. The second part infers that there are no unmeasured baseline variables confounding the relationships between the treatment, the mediator, and the outcome. This is potentially satisfied by controlling for all baseline covariates potentially confounding these relationships. The magnitude of departure from the second assumption is estimated by a sensitivity analysis assessing the contribution of unmeasured confounding variables (Fig 1).²⁸ 

The causal mediation analysis yields average causally mediated, direct, and total effects as well as the proportion of the mediated effects (of the total effect) as point estimates with 95% confidence intervals (CIs) and the associated P values. For the estimation of the mediation effects, a lack of significant interaction between the trial arms is assumed (ie, the mediation effects are independent of the trial arm and averaged over the entire study population). We tested the effect of relaxing this assumption by identifying the outcomes for which the average causal effects were significantly different between the trial arms.

With 400 children enrolled, the CHAT study was powered to detect differences in the primary outcome between children grouped by the type of treatment.⁸  For tests of relationships between each arm of the mediation analysis, an individual effect size of 0.30 (medium) as measured by partial correlation is assumed (treatment to mediator, mediator to outcome, and treatment to outcome; Fig 1).³¹  A sample size of 400 and type 1 error rate of 5% yield a power of nearly 1 for the mediation analysis. Additionally, 2 types of sensitivity analysis were performed; the first was performed to account for the effects of missing values and the second to determine the possible existence of unobserved pretreatment covariates when significant causal mediation effects were identified. The methodologic details of the 2 sensitivity analyses are described in the Supplemental Information. P < .05 was considered significant throughout.

The primary results of the CHAT study have been published.⁸  Of the 464 children enrolled in the study and who underwent random assignment, 398 were included in the current study on the basis of the availability of complete polysomnographic data. Supplemental Fig 2 reveals patient flow into the current study from the CHAT. The baseline characteristics of children in the study are shown in Table 2. By using the criteria defined in the CHAT, a significantly greater number of children (244 [61%]) had polysomnographic resolution of OSA compared with those who did not experience resolution (154 [39%]; P < .001). Of those without polysomnographic resolution, 96% had an AHI ≥2 and 50% had an AHI ≥5 on follow-up polysomnography. In this group, there were a significantly greater proportion of African American children (63% vs 47%; P = .001) and obese children (47% vs 25%; P < .001) compared with the group without resolution. The mean baseline AHI in the group with complete resolution (5.8 [95% CI, 5.1 to 6.4]) was significantly lower than the group without resolution (8.3 [95% CI, 7.4 to 9.2]; P < .001).

Early adenotonsillectomy resulted in significantly greater changes in 12 out of 18 outcomes compared to watchful waiting (Table 3). Of these, the largest effect sizes were observed for the changes in symptoms measured by the PSQ-SRBD (Cohen’s d, 1.35; P < .001) and disease-specific quality of life assessed by OSA-18 (Cohen’s d, 0.94; P < .001).

Table 4 shows the results of mediation analysis averaged over both arms of the trial when polysomnographic resolution of OSA was considered the mediator. The total effects of treatment, defined as the sum of mediation and direct effects, were significant for 12 of the 18 outcomes. Small but statistically significant average mediation effects were identified for the PSQ-SRBD (proportion mediated: 0.13; 95% CI, 0.07 to 0.21; P < .001) and OSA-18 (proportion mediated: 0.11 [95% CI, 0.04 to 0.20]; P = .004). Similarly, the change in AHI between baseline and follow-up revealed significant mediation effects (Table 5) for the PSQ-SRBD (proportion mediated: 0.18 [95% CI, 0.11 to 0.26]; P < .001) and OSA-18 (proportion mediated: 0.20 [95% CI, 0.10 to 0.31]; P < .001). However, 16 out of 18 outcomes were not associated with significant contribution from either mediator. When OSA resolution was defined in a more robust fashion as an AHI <1.5, OAI <1, and absence of hypoxemia as measured by oximetry (Supplemental Table 6), the mediation effects were limited to the PSQ-SRBD (proportion mediated: 0.06 [95% CI, 0.02 to 0.12]; P = .005). Investigation of interaction between the trial arms revealed that polysomnographic resolution in children undergoing surgery mediated weight gain as opposed to weight loss in children undergoing watchful waiting (difference in mediation effect: 2.02 [95% CI, 0.40 to 3.85]; P = .01; Supplemental Table 7 for OSA resolution as mediator and Supplemental Tables 8 and 9 for other mediators). None of the other 17 outcomes revealed statistically significant interaction effects.

The first sensitivity analysis revealed that the missing observations from the data did not significantly alter causal inferences (data not shown). Supplemental Table 10 and Supplemental Figs 3–5 represent sensitivity analyses for the significant mediation effects identified for 3 separate causal pathways. In summary, sequential ignorability was less likely to be violated for the mediation effects in the PSQ-SRBD (sensitivity parameter of −0.40 for the watchful waiting group and −0.30 for the adenotonsillectomy group) compared to the OSA-18 (sensitivity parameter of −0.30 and −0.10, respectively, for the 2 arms of the trial).

In this study, polysomnographic resolution and changes in severity of OSA accounted for small but significant proportions of changes in both symptoms and disease-specific quality of life in children treated for OSA. Importantly, polysomnographic resolution of OSA or changes in its severity did not causally impact 16 out of the 18 treatment outcomes. These results highlight the limited utility associated with the use of polysomnographic thresholds in the management of OSA.

The principal outcome reported in the treatment of pediatric OSA is the resolution of the condition by polysomnography. This approach is supported by several nonrandomized studies with both prospective³²  and retrospective³³  designs as well as multiple meta-analyses.^34–36  As shown by the CHAT study, children who were African American, obese, or with greater baseline severity of OSA were more likely to have persistent OSA at follow-up regardless of the treatment arm.⁸ 

Additionally, the comparison of treatment effects between the 2 arms of the trial also revealed statistically significant increases in weight gain and HOMA-IR in the surgical arm. Although weight gain after adenotonsillectomy has been described previously,³⁷  the effect sizes of these differences were small in the current study.

Because of the risk of persistent OSA in 10% to 50% of children undergoing adenotonsillectomy, postoperative polysomnography is recommended²  for early detection and further treatment.¹  On the basis of this recommendation, in the current study, 50% of the children (77 out of 154) without resolution of OSA may have been potential candidates for further treatment ranging from additional surgery to continuous positive air pressure therapy solely on the basis of an AHI exceeding 5 at follow-up.

The CHAT study remains the only randomized trial to this date used to investigate the benefits of surgery over watchful waiting for outcomes related to childhood OSA. The trial additionally provided the opportunity to examine the isolated effects of changes in polysomnographic severity of OSA or resolution on its treatment. Importantly, we show that the results of the causal mediation analysis suggest that a majority of the treatment-related changes in outcomes of OSA in children are not attributable to polysomnographic changes in its severity or resolution. Given that 12 out of the 18 outcomes in the trial revealed significant average changes with treatment, alterations in other polysomnographic parameters such as sleep architecture may play a mediating role and merit further investigation.

The use of polysomnography is central to the guidelines from the American Academy of Pediatrics,¹  American Academy of Sleep Medicine,²  American Academy of Otolaryngology–Head and Neck Surgery,³  and the ERS⁴  for the diagnosis and management of OSA, including the 500 000 children undergoing adenotonsillectomy each year. The ERS suggests treating a child with an AHI >5 even in the absence of comorbidities.⁴  Such an approach increases the risk of avoidable surgical morbidity in children considered candidates for further treatment solely on the basis of empirical polysomnographic thresholds. Moreover, a stepwise approach that seeks complete resolution of OSA in children has been proposed with a recommendation to measure outcomes on the basis of polysomnography at each stage.⁴  Liberal polysomnography as proposed in these guidelines may not be justifiable given the costs and resources needed.^7,38 

The principal strengths of this study are related to the robust criteria used to define inclusion and exclusion of subjects in the only randomized trial to assess the benefit of adenotonsillectomy over watchful waiting for the treatment of childhood OSA. The causal mediation analysis described in the current study has been promulgated to provide mechanistic explanations for outcomes related to interventions.³⁹  Furthermore, the causal models described here are generally agnostic to the nature of the underlying data distributions.²⁸  Children were enrolled from multiple sites across the United States, underscoring the generalizability of the results and bias mitigation efforts. The 18 outcomes examined as part of the trial were derived from all possible domains potentially impacted by OSA in children. The weaknesses of the study are similar to those listed by the original trial, which include the narrow range of age of children, the exclusion of children with an AHI exceeding 30 and a relatively short period of follow-up. Additionally, identification of other causal mediators related to the treatment of OSA is a subject of future investigation.

The mediation analysis of the results obtained from the CHAT study suggests a limited role for polysomnographic resolution or changes in severity of OSA in causally influencing the outcomes related to its treatment. These results caution against the use of empirical AHI-based thresholds to assess outcomes related to treatment. Further studies are necessary to identify possible mediators of outcomes of pediatric OSA to better define the severity of the condition and reduce the cost of diagnosis.

We thank the University of Maryland Biostatistics Core of the Institute for Clinical and Translational Research for statistical consultations. We also acknowledge Drs Abraham Kanate, Anita Shet, and Arun Shet for their useful feedback on earlier versions of the article.

AHI

apnea hypopnea index

BRIEF

Behavior Rating Inventory of Executive Function

CBCL

Child Behavior Checklist

CHAT

Childhood Adenotonsillectomy Trial

CI

confidence interval

CRP

C-reactive protein

ERS

European Respiratory Society

ESS

Epworth Sleepiness Scale

HOMA-IR

Homeostasis Model Assessment for Insulin Resistance

NEPSY

Developmental Neuropsychological Assessment

OAI

obstructive apnea index

OSA

obstructive sleep apnea

OSA-18

Obstructive Sleep Apnea-18

PedsQL

Pediatric Quality of Life Inventory

PSQ-SRBD

Pediatric Sleep Questionnaire Sleep-Related Breathing Disorder