Medical Student Emory University Department of Orthopaedics
Disclosure(s):
Kevin Y. Heo, BS: No financial relationships to disclose
Introduction: Surgical site infection (SSI) following spine surgery can result in significant morbidity and mortality. Compared to anterior cervical procedures, posterior approaches have consistently demonstrated higher rates of SSI. However, relatively few studies have incorporated predictive analytics to identify patients at high risk for SSI following posterior cervical spine surgery. Therefore, the objective of the study was to utilize machine learning (ML) models to assess risk factors for SSI after posterior cervical spine surgery.
Methods: Data was retrospectively collected from a national database between 2009 and 2019 for patients who underwent posterior cervical spine surgery. SSI incidence was captured 30-days after surgery. Two different methodologies were used to identify SSI risk: 1) a novel predictive model derived from multivariable logistic regression, and 2) a tree-based extreme-gradient boosting (XGBoost) model. These two methods were compared utilizing area-under-the-curve (AUC) statistics.
Results: Among the 53,367 patients identified, the overall rate of SSI was 2.9%, with the highest proportion in those whose primary billing code was associated with posterior cervical instrumented fusion (4.2%). The novel predictive model exhibited an AUC of 0.60, identifying patients with hemiplegia, morbid obesity (BMI>40), alcohol use disorder, hypertension, and depression as being most predictive of 30-day SSI. The XGBoost model had an AUC of 0.71, with the five clinical variables found to be most predictive of SSI including: active tobacco use, peripheral vascular disease, hypertension, hemiplegia, and diabetes. Based on the XGBoost model, patients who did not have any of these 5 key risk factors had an SSI rate of 2.3%, while patients who had one of the risk factors had an SSI rate of 3.4% (XGBoost odds ratio [OR] 1.49, P=0.002). Subsequently, patients with increasing number of risk factors had higher risks of SSI.
Conclusion : Two different predictive models identified key variables that are most closely associated with SSI over 30-days after posterior cervical spinal surgery. It is important to note that other factors not available in the database, including randomness, may have accounted for the variance of SSI onset. Nevertheless, these findings can be leveraged to help surgeons stratify their patients for SSI risk, and potentially guide postoperative medical management.