Evaluating parsimonious risk-adjustment models for comparing hospital outcomes with vascular surgery

Abstract

Most outcomes registries use a large number of variables to control for differences in patients. We sought to determine whether fewer variables could be used for risk adjustment without compromising hospital quality comparisons. We used prospective, clinical data from the American College of Surgeons-National Surgical Quality Improvement Program (ACS-NSQIP) for five commonly performed inpatient vascular procedures (N = 24,744). For each of the five operations, we compared the ability of two parsimonious models (an intermediate model, using the top five variables for each procedure and a limited model using the top 2 variables from each procedure) and the full model (up to 42 variables) to predict the risk of mortality and morbidity at the patient and hospital level. The parsimonious model was similar to the full model in all comparisons. For the five procedures, the intermediate, limited, and full models all had very similar discrimination at the patient-level (C indices of 0.87 vs 0.85 vs 0.87 for mortality and 0.77 vs 0.75 vs 0.77 for morbidity), and similar calibration, as assessed with the Hosmer-Lemeshow test. In evaluating hospital-level morbidity and mortality rates, the correlations between the parsimonious and full models were very high for both mortality (>0.97 across operations) and morbidity (>0.97 across operations). Hospital quality comparisons for vascular surgery can be adequately risk-adjusted using a small number of important variables. Reducing the number of variables collected will significantly decrease the burden of data collection for hospitals choosing to participate in the vascular module of the ACS-NSQIP.