Study design and data source
We performed a nationwide observational study to determine whether patient outcomes improve in hospitals with a significantly increased volume of high-risk surgery over time and whether a learning effect existed at the individual hospital level. We first defined three groups of hospitals according to the trend of the volume of surgical procedures over a 5-year period, that is, for a given hospital, the volume of a specific procedure increased, decreased, or did not change. Second, we compared the average patient outcomes and their evolution over time between these three defined hospital groups, taking into account potential confounding factors related to hospital and patient characteristics. To test the robustness of our results, we repeated this scheme across ten high-risk surgical and interventional procedures in various specialties and considered different patient outcomes.
This study used the French Medical Information System (Programme de Médicalisation des Systèmes d’Information [PMSI]), which is a large acute rate hospital database with prospectively collected data from all public and private hospitals in France. The database is routinely implemented for the purpose of care reimbursement, which in turn led to strong accuracy and exhaustive collection of data. Thus, no patients were assumed to be lost to follow-up during the study period. Moreover, the PMSI has a system of coding with strict variable definitions and a subset of records audited on a regular basis to avoid coding errors. Inpatient stays are converted into one Diagnosis-Related Group based on standard discharge abstracts containing compulsory information about the patient, primary and secondary diagnoses using the International Classification of Diseases (10th revision - ICD-10 codes), emergency status, and procedural codes associated with the care provided using a detailed classification.
From the PMSI database, we extracted data on patient demographics, co-morbidities according to the Elixhauser algorithm [27], the type and emergency context of the procedure, and discharge by transfer to another acute care hospital. We also characterized each hospital according to its status (i.e., teaching, public, or private for-profit), degree of specialization (i.e., proportion of admissions logged for each studied procedure in the related surgical department), and attraction rate (i.e., the proportion of patients living in another geographical area than that of the hospital location where they underwent each studied procedure). To define patients’ socioeconomic status, we extracted the median income of the patients’ residence provided by the National Institute of Statistics and Economic Studies.
This study was approved by the National Data Protection Commission (Commission Nationale Informatique et Libertés) in accordance with French ethical directives and was registered on clinicaltrial.gov (NCT02788331).
Study population and outcomes
We included all patients who underwent one of the following ten procedures from January 1, 2010 to December 31, 2014: resection of a digestive cancer (i.e., colectomy, proctectomy, esophagectomy, gastrectomy, and pancreatectomy), intervention on the cardiovascular system (i.e., percutaneous coronary intervention [PCI]), coronary-artery bypass grafting [CABG], carotid endarterectomy, and elective repair of abdominal aortic aneurysm [AAA]), and urgent hip fracture repair (Appendix S1). The choice to focus on those procedures was guided by available evidence suggesting the existence of volume-outcome relationships based on cross-sectional studies [2–5]. Each procedure was identified from the PMSI database by combining specific diagnoses and procedural codes.
For each studied procedure, all patients from hospitals not performing at least one procedure per year were removed from the dataset. Furthermore, patients <18 years old, who experienced ambulatory care, or with data inaccuracies were excluded. After the washout for every procedure since 2009, we only selected the first hospitalization of each patient identified as the index stay (except in the case of hip fracture, in which two stays were potentially included if the second stay occurred at least 30 days after the first discharge), using unique, anonymous patient numbers that linked all his/her stays in acute care.
The following patient outcomes were analysed: in-hospital mortality, reoperation, and potentially avoidable hospital readmission. In-hospital mortality and reoperation were defined as death and reoperation, respectively, within a maximum of 30 days post-procedure, whereas potentially avoidable readmission was studied within 30 days of the index stay discharge [28, 29].
Statistical analysis
To classify hospitals based on their volume change over time, we calculated individual hospital volume for each of the ten studied procedures as the total number of patients treated by each hospital within each year. Subsequently, hospitals were divided into three groups based on whether their annual procedure volumes were increasing, decreasing or remaining stable over a 5-year period. We used the random slopes of multilevel Poisson models, taking into account the annual repeated measures of hospital volume for each procedure. These slopes were categorized into three groups using the K-means method to avoid arbitrary determination of thresholds and to account for intra-group variances that could vary [30].
For each procedure, to determine if mortality was altered in patients admitted to hospitals with significantly increased volume changes over time and if a learning effect existed at the individual hospital level, we used cox regressions, taking into account the clustering effect of patients within hospitals with robust variance estimator (i.e., patients treated and outcomes within a particular hospital tended to be more similar than those in another hospital), the follow-up that varied from one patient to another, and the hospital discharge that represented a censure of outcome [31, 32]. Furthermore, hospital learning effect was investigated by examining the interaction between hospital groups and year of procedure. To adjust mortality for case mix variations, we considered patient (age, gender, Elixhauser list of comorbidities, type and year of procedure, transfer, emergency admission, and median income) and hospital (hospital status, volume of procedures, specialization degree, and attraction rate) characteristics. Restricted cubic splines were used for continuous variables in the adjustment scheme [33].
To test the robustness of our results, we repeated this analysis across secondary outcomes (reoperation and unplanned hospital readmission) using Fine and Gray’s models to consider the, competing risk of death. Model estimates were presented as adjusted hazards ratios (HR) with corresponding 95% confidence interval (95% CI). Data manipulation and analyses were performed using SAS version 9.4 (SAS Institute Inc, Cary, NC) and R version 3.2.1 (R Foundation for Statistical Computing, Vienna, Austria) software.