Failure of Colorectal Surgical Site Infection Predictive Models Applied to an Independent Dataset: Do They Add Value or Just Confusion?

Background Colorectal surgical site infections (C-SSIs) are a major source of postoperative morbidity. Institutional C-SSI rates are modeled and scrutinized, and there is increasing movement in the direction of public reporting. External validation of C-SSI risk prediction models is lacking. Factors governing C-SSI occurrence are complicated and multifactorial. We hypothesized that existing C-SSI prediction models have limited ability to accurately predict C-SSI in independent data. Study Design Colorectal resections identified from our institutional ACS-NSQIP dataset (2006 to 2014) were reviewed. The primary outcome was any C-SSI according to the ACS-NSQIP definition. Emergency cases were excluded. Published C-SSI risk scores: the National Nosocomial Infection Surveillance (NNIS), Contamination, Obesity, Laparotomy, and American Society of Anesthesiologists (ASA) class (COLA), Preventie Ziekenhuisinfecties door Surveillance (PREZIES), and NSQIP-based models were compared with receiver operating characteristic (ROC) analysis to evaluate discriminatory quality. Results There were 2,376 cases included, with an overall C-SSI rate of 9% (213 cases). None of the models produced reliable and high quality C-SSI predictions. For any C-SSI, the NNIS c-index was 0.57 vs 0.61 for COLA, 0.58 for PREZIES, and 0.62 for NSQIP: all well below the minimum "reasonably" predictive c-index of 0.7. Predictions for superficial, deep, and organ space SSI were similarly poor. Conclusions Published C-SSI risk prediction models do not accurately predict C-SSI in our independent institutional dataset. Application of externally developed prediction models to any individual practice must be validated or modified to account for institution and case-mix specific factors. This questions the validity of using externally or nationally developed models for "expected" outcomes and interhospital comparisons.