Development and Validation of Perioperative Risk-Adjustment Models for Hip Fracture Repair, Total Hip Arthroplasty, and Total Knee Arthroplasty

BACKGROUND: Comparing outcomes across providers requires risk-adjustment models that account for differences in case mix. The burden of data collection from the clinical record can make risk-adjusted outcomes difficult to measure. The purpose of this study was to develop risk-adjustment models for hip fracture repair (HFR), total hip arthroplasty (THA), and total knee arthroplasty (TKA) that weigh adequacy of risk adjustment against data-collection burden. METHODS: We used data from the American College of Surgeons National Surgical Quality Improvement Program to create derivation cohorts for HFR (n = 7000), THA (n = 17,336), and TKA (n = 28,661). We developed logistic regression models for each procedure using age, sex, American Society of Anesthesiologists (ASA) physical status classification, comorbidities, laboratory values, and vital signs-based comorbidities as covariates, and validated the models with use of data from 2012. RESULTS: The derivation models' C-statistics for mortality were 80%, 81%, 75%, and 92% and for adverse events were 68%, 68%, 60%, and 70% for HFR, THA, TKA, and combined procedure cohorts. Age, sex, and ASA classification accounted for a large share of the explained variation in mortality (50%, 58%, 70%, and 67%) and adverse events (43%, 45%, 46%, and 68%). For THA and TKA, these three variables were nearly as predictive as models utilizing all covariates. HFR model discrimination improved with the addition of comorbidities and laboratory values; among the important covariates were functional status, low albumin, high creatinine, disseminated cancer, dyspnea, and body mass index. Model performance was similar in validation cohorts. CONCLUSIONS: Risk-adjustment models using data from health records demonstrated good discrimination and calibration for HFR, THA, and TKA. It is possible to provide adequate risk adjustment using only the most predictive variables commonly available within the clinical record. This finding helps to inform the trade-off between model performance and data-collection burden as well as the need to define priorities for data capture from electronic health records. These models can be used to make fair comparisons of outcome measures intended to characterize provider quality of care for value-based-purchasing and registry initiatives.