Comparison of postoperative complication risk prediction approaches based on factors known preoperatively to surgeons versus patients

Abstract

Background: Estimating the risk of postoperative complications can be performed by surgeons with detailed clinical information or by patients with limited information. Our objective was to compare three estimation models: (1) the All Information Model, using variables available from the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP); (2) the Surgeon Assessment Model, using variables available to surgeons preoperatively, and (3) the Patient-Entered Model, using information that patients know about their own health.

Study design: Using the ACS NSQIP 2011 data for general and colon surgery, standard ACS NSQIP regression methods were used to develop models. Each model examined Overall and Serious Morbidity as outcomes. The models were assessed using the c-statistic, Hosmer-Lemshow statistic, and Akaike Information Criterion.

Results: The overall morbidity rate was 13.0%, and the serious morbidity rate was 10.5% for patients undergoing general surgery (colon surgery: 31.8% and 26.0%, respectively). There was a small decrement in the c-statistic as the number of predictors decreased. The Akaike Information Criterion likelihood ratio increased between the All Information and Surgeon Assessment models, but decreased in the Patient-Entered Model. The Hosmer-Lemshow statistic suggested good model fit for five colon surgery models and one general surgery model.

Conclusion: Although a small decline in model performance was observed, the magnitude suggests that it may not be clinically meaningful as the risk predictions offered are superior to simply providing unadjusted complications rates. The Surgeon Assessment and Patient-Entered models with fewer predictors can be used with relative confidence to predict a patient's risk.