Surgical specialty and preoperative medical consultation based on commercial health insurance claims

We evaluated a retrospective cohort of surgical claims from Premera Blue Cross, a commercial insurance company based in Seattle, Washington (WA). All patients were Washington State residents, aged 18–80 years old, and had undergone one of 30 a priori selected common inpatient or outpatient surgical procedures in calendar year 2010. We selected surgical procedures to include commonly performed procedures representing a spectrum of surgical risk (see the “Appendix” section for included procedures and their frequencies in the study cohort). We included eight different surgical specialties: general surgery, ophthalmology, gynecology, orthopedics, urology, otolaryngology, neurosurgery, and vascular surgery. Surgical events were nested within surgeons, and surgeons were nested within provider organizations.

We identified eligible procedures by the patient’s first occurrence of a Current Procedural Terminology (CPT) code for one of the specified procedures. We identified the primary endpoint—preoperative medical consultation within the 42-day preoperative period—by the presence of any CPT code for outpatient consultation (99242-99245) or inpatient consultation (99252-99255). We included consultations provided by family physicians, general internists, pulmonologists, cardiologists, endocrinologists, or nurse practitioners. We also included office visits modified with an International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) code v72.81-v72.84, indicating that the visit was a separately billed preoperative evaluation. We chose a period of 42 days preceding surgery to maximize the ability to capture preoperative consultations that may have led to a delay in scheduling the surgical procedure.

We calculated the Deyo comorbidity index using diagnostic codes present within 365 days before surgery (Deyo et al. 1992). The Deyo comorbidity index was used as an ordinal variable with categories 0, 1, 2, and ≥ 3. We also calculated the Revised Cardiac Risk Index (RCRI) (Lee et al. 1999), which predicts cardiac complications. However, we did not include the RCRI in the final analyses because very few patients were class III or IV risk (98.6% were class I or II), and the RCRI index did not significantly predict consultation after adjusting for the Deyo index. Surgical procedures were categorized into low, intermediate, and high levels of surgical risk. Patient demographics included age and gender. We used the ZIP code of patient residence to allocate patients to seven hospital referral regions (HRRs) (Everett, Olympia, Seattle, Spokane, Tacoma, and Yakima (all WA cities), and Portland, OR) and to determine patient’s Rural Urban Commuting Area (RUCA) category (urban, large rural city, small rural town, and isolated rural town). HRRs are approximate health care markets for tertiary care with at least one hospital that performs major cardiovascular and neurosurgery procedures (Dartmouth Medical School and Center for the Evaluative Clinical Sciences 1996). RUCA codes are a classification scheme for characterizing all of US ZIP codes according to their urban or rural character (WWAMI Rural Health Research Center n.d.).

We summarized counts of consultation and we ordered them by day preceding surgery up to 42 days before the index surgery. We conducted bivariate analyses to compare characteristics of patients who did or did not undergo preoperative consultation. We estimated the association between preoperative consultation (outcome variable) and surgical specialty (explanatory variable of interest) by fitting a log-Poisson generalized estimating equation (GEE) regression model with robust standard errors. GEE permits obtaining marginal population-averaged interpretations of hierarchical data using non-linear link functions. The exponents of beta coefficients from a log-Poisson model are true risk ratios in contrast to the odds ratios produced by a logistic regression. Odds ratios approximate risk ratios but are inflated (i.e., overstate effects) when outcomes are more common or as effects become larger (Kleinman and Norton 2009). Since surgeons are nested within provider organizations without any non-nested clusters, we clustered on provider organization, and the robust standard errors account for clustering at both the provider and organization levels (Betensky et al. 2000). We adjust for potential confounders: age, gender, Deyo comorbidity index, surgical risk, HRRs, and RUCA categories. The log-Poisson model passed model diagnostic tests: the Hosmer-Lemeshow statistic for model misspecification, the Pearson correlation of the raw-scale residuals and predictions, and Pregibon’s link test. A two-sided α level of 0.05 was considered for statistical significance. We report effects as relative risks in the text and Table  2 and also report selected absolute adjusted model estimates in the text and Fig.  2 (main effects model only), as specified (hereafter , adjusted likelihoods). All statistical analyses were performed using Stata 14 (Stata Corporation, College Station, TX, USA).

This study is based on private insurance claims, which may be affected by coding errors or misclassification of preoperative consultation. Consultations most commonly occurred on preoperative days 7 and 14, and with a very low occurrence of visits preceding preoperative day 30, which suggests that most of these visits were associated with a planned surgery (i.e., low misclassification of unrelated visits as preoperative consultations). Furthermore, this pattern’s consistency with prior findings (including our study of an integrated health care system) corroborates the construct validity of our measure of preoperative consultations.

If preoperative medical consultations delayed or canceled surgeries, our data would not include those consultations. However, we believe this limitation is minor, as we used a relatively long, 42-day preoperative window. It is also unknown whether the information resulting from these consultations altered the patient perioperative management. Nevertheless, we have demonstrated that there is a substantial usage of preoperative medical consultation also for patients with limited or no comorbidities, many of whom are undergoing low-risk surgeries. Although this sample is relatively large for studies of preoperative consultation in the US, the sample still had limited power to detect small effects. The study sample included a small number of vascular surgery and neurosurgery patients, and this limits our ability to draw conclusions specific to these specialties regarding consultations. In the absence of a broad consensus regarding an optimal surgical-risk categorization for all procedures listed, we arbitrarily assigned surgical risk based on similarities with other known procedures in the same categories, and based on clinical judgment. This risk categorization imperfectly differentiates average risk across procedures. We included other covariates to additionally adjust for individual- and geographic-specific factors that predict preoperative consultations, although we could not entirely eliminate selection bias in this observational study.