Improving surgical site infection risk-adjustment methods for hip and knee arthroplasties: data from the National Healthcare Safety Network

Background: Nearly 1 million hip and knee arthroplasty procedures are performed annually in the US.  Feedback of risk-adjusted surgical site infection (SSI) rates to healthcare facilities is effective in reducing SSI.  The National Healthcare Safety Network (NHSN) currently stratifies SSI rates by the National Nosocomial Infection Surveillance (NNIS) risk index, a score based on procedure duration, wound class, and American Society of Anesthesiology (ASA) score.  Using more variables and flexibility in parameterization may improve risk-adjustment and increase the relevance of SSI rate comparisons to the surgical community.

Objective: To determine whether a predictive modeling approach using routinely collected NHSN variables can improve hip and knee arthroplasty SSI risk-adjustment.

Methods: We analyzed data on all inpatient hip (n=131,948) and knee (n=172,591) arthroplasties reported to NHSN for 2006-2008 and used stepwise logistic regression to construct models predicting expected SSI rates based on patient, procedure, and facility characteristics.  The models’ ability to predict SSI events was compared to the NNIS risk index using area under the receiver-operator characteristic curves (AUC).

Results: A total of 1,654 hip arthroplasty SSI (12.7 per 1,000 procedures) were reported to NHSN during the study period: 81% were identified at readmission, 60% were caused by S. aureus, and 22% were joint infections.  The final predictive model for hip arthroplasties included ASA categories (1/2 vs. 3 vs. 4/5); patient age >80 years; duration of procedure; whether the procedure was a hemi-arthroplasty, a revision, performed due to trauma, or performed under general anesthesia; and facility bed size.  The AUC for this model was 7% higher than for the NNIS index, indicating improved ability to predict SSI.  A total of 1,547 knee arthroplasty SSI (9.0 per 1,000 procedures) were reported: 82% were identified at readmission, 55% were caused by S. aureus, and 30% were joint infections.  The final predictive model for knee arthroplasties included ASA categories (1/2 vs. 3 vs. 4/5); patient age; male gender; duration of procedure; whether the procedure was a revision or performed due to trauma; and facility medical school affiliation.  The AUC for this model was 6% higher than for the NNIS index, also indicating improved ability to predict SSI.

Conclusions: Predictive modeling methods led to improved SSI prediction compared to current methods used in NHSN.  Expected facility-specific SSI rates generated from these models can be compared to observed rates using standardized infection ratios to gauge infection control performance.  Next steps include evaluating the impact that improved risk-adjustment has on the relevance of these data to the surgical community and the ability of this new method to drive improvements in surgical practice.