Risk Prediction Accuracy Differs for Emergency Versus Elective Cases in the ACS-NSQIP.

Abstract

Accurate risk estimation is essential when benchmarking surgical outcomes for reimbursement and engaging in shared decision-making. The greater complexity of emergency surgery patients may bias outcome comparisons between elective and emergency cases. To test whether an established risk modelling tool, the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) predicts mortality comparably for emergency and elective cases. From the ACS-NSQIP 2011-2012 patient user files, we selected core emergency surgical cases also common to elective scenarios (gastrointestinal, vascular, and hepato-biliary-pancreatic). After matching strategy for Common Procedure Terminology (CPT) and year, we compared the accuracy of ACS-NSQIP predicted mortality probabilities using the observed-to-expected ratio (O:E), c-statistic, and Brier score. In all, 56,942 emergency and 136,311 elective patients were identified as having a common CPT and year. Using a 1:1 matched sample of 37,154 emergency and elective patients, the O:E ratios generated by ACS-NSQIP models differ significantly between the emergency [O:E = 1.031; 95% confidence interval (CI) = 1.028-1.033] and elective populations (O:E = 0.79; 95% CI = 0.77-0.80, P < 0.0001) and the c-statistics differed significantly (emergency c-statistic = 0.927; 95% CI = 0.921-0.932 and elective c-statistic = 0.887; 95% CI = 0.861-0.912, P = 0.003). The Brier score, tested across a range of mortality rates, did not differ significantly for samples with mortality rates of 6.5% and 9% (eg, emergency Brier score = 0.058; 95% CI = 0.048-0.069 versus elective Brier score = 0.057; 95% CI = 0.044-0.07, P = 0.87, among 2217 patients with 6.5% mortality). When the mortality rate was low (1.7%), Brier scores differed significantly (emergency 0.034; 95% CI = 0.027-0.041 versus elective 0.016; 95% CI = 0.009-0.023, P value for difference 0.0005). ACS-NSQIP risk estimates used for benchmarking and shared decision-making appear to differ between emergency and elective populations.