Postoperative Hypoxemia in Orthopedic Patients with Obstructive Sleep Apnea

Abstract

Criteria to determine which patients with obstructive sleep apnea (OSA) require intensive postoperative monitoring are lacking. Our postoperative OSA patients are all intensively monitored in the PACU and can provide such data. Thus, we reviewed patient records to determine incidence and risk factors for postoperative hypoxemia in OSA patients and subsequent association with postoperative complications. Five hundred twenty-seven charts of patients with OSA based on preoperative ICD-9 codes were reviewed for outcomes including episodes of hypoxemia and hypercarbia. Univariate analysis, logistic regression, and propensity analysis were performed to determine independent risk factors for hypoxemia and association with adverse outcomes. Thirty-three and 11 percent of these patients developed hypoxemia or hypercarbia. Risk factors for hypoxemia were hypercarbia, home bronchodilator use, BMI ≥35, and estimated blood loss ≥250 ml. Patients with hypoxemia had significantly more respiratory interventions and increased intensity of care. Patients with hypoxemia had significantly increased length of stay and risk of wound infections. Severe hypoxemia was associated with significantly more interventions than mild hypoxemia. Propensity analysis confirmed significant association of hypoxemia with adverse outcomes after adjustment for pre-existing risk factors. We conclude that postoperative hypoxemia in OSA patients is associated with adverse outcomes. Risk factors for hypoxemia were identified to guide allocation of monitoring resources to high-risk patients.

Appendix: specifics of logistic regression

All statistical analyses were performed using SAS version 9.1.3 (SAS Institute, Cary, NC, USA). Model fitting for multivariate logistic regression started with a full model including all risk factors that were significant in univariate analysis. The full model consisted of an outcome variable of hypoxemia, five dichotomous (BMI ≥35, episode of hypercarbia, use of home bronchodilators, ASA class ≥3, and general versus regional anesthesia), and three continuous (IV fluids, estimated blood loss, and length of procedure) predictors and was adjusted for type of surgical procedure (spine, hip or knee replacement, lower extremity, and upper extremity). Continuous variables were tested for continuity and linearity and were virtually linear in a graphic analysis [9]. For each individual predictor, odds ratio, 95% confidence interval, and p value were computed. Multicollinearity was judged by checking the value inflation factor and condition index, and standard criterion for absence of collinearity was sufficiently met (value inflation factor<10 and condition index <30) [2]. Multivariate logistic regression models were fitted and selected using backward elimination, and models with the lowest Akaike Information Criterion were selected [1]. The Hosmer–Lemeshow test of goodness-of-fit was performed and indicated that the final logistic model fit the data well (p = 0.43) [1].

We also checked data for patient clustering effect by anesthesiologist or surgeon (e.g., whether a single subgroup of anesthesiologists or surgeons was involved with most patients). Chi square tests were conducted to compare the distributions of patient within each cluster (anesthesiologists or surgeons), and no significant association was found between hypoxemia and anesthesiologist (p = 0.33) and surgeon (p = 0.65). Furthermore, the generalized estimating equations (GEE) approach for logistic regression was also applied to identify risk factors for hypoxemia. Unlike standard logistic regression, GEE logistic regression allows for dependence within clusters, specifically clustering by anesthesiologist or surgeon in this case [13]. The same methods for model fitting and model selection for logistic regression were also applied to GEE. After analysis, both GEE logistic and standard logistic final models contained the same set of independent risk factors in Table 5 with nearly identical coefficient estimates. As neither the Chi square tests nor the GEE logistic regression indicated a clustering effect, results from standard logistic regression were used.