Performance Of Surgical Site Infection Risk Prediction Models In Colorectal Surgery: External Validity Assessment From Three European National Surveillance Networks

Objective: To assess the validity of multivariable models for predicting risk of surgical site infection (SSI) after colorectal surgery based on routinely collected data in national surveillance networks. Design: Retrospective analysis performed on 3 validation cohorts. Patients: Colorectal surgery patients in Switzerland, France, and England, 2007-2017. Methods: We determined calibration and discrimination (ie, area under the curve, AUC) of the COLA (contamination class, obesity, laparoscopy, American Society of Anesthesiologists [ASA]) multivariable risk model and the National Healthcare Safety Network (NHSN) multivariable risk model in each cohort. A new score was constructed based on multivariable analysis of the Swiss cohort following colorectal surgery, then based on colon and rectal surgery separately. Results: We included 40,813 patients who had undergone elective or emergency colorectal surgery to validate the COLA score, 45,216 patients to validate the NHSN colon and rectal surgery risk models, and 46,320 patients in the construction of a new predictive model. The COLA score's predictive ability was poor, with AUC values of 0.64 (95% confidence interval [CI], 0.63-0.65), 0.62 (95% CI, 0.58-0.67), 0.60 (95% CI, 0.58-0.61) in the Swiss, French, and English cohorts, respectively. The NHSN colon-specific model (AUC, 0.61; 95% CI, 0.61-0.62) and the rectal surgery-specific model (AUC, 0.57; 95% CI, 0.53-0.61) showed limited predictive ability. The new predictive score showed poor predictive accuracy for colorectal surgery overall (AUC, 0.65; 95% CI, 0.64-0.66), for colon surgery (AUC, 0.65; 95% CI, 0.65-0.66), and for rectal surgery (AUC, 0.63; 95% CI, 0.60-0.66). Conclusion: Models based on routinely collected data in SSI surveillance networks poorly predict individual risk of SSI following colorectal surgery. Further models that include other more predictive variables could be developed and validated.