Construct validation of machine learning for accurately predicting the risk of postoperative surgical site infection following spine surgery

Abstract

Background: This study aimed to evaluate the risk factors for machine learning (ML) algorithms in predicting postoperative surgical site infection (SSI) following spine surgery. Methods: This prospective cohort study included 986 patients who underwent spine surgery at Taizhou People's Hospital Affiliated to Nanjing Medical University from January 2015 to October 2022. Supervised ML algorithms included support vector machine, logistic regression, random forest, XGboost, decision tree, k-nearest neighbour, and naïve Bayes (NB), which were tested and trained to develop a predicting model. The ML model performance was evaluated from the test dataset. We gradually analysed their accuracy, sensitivity, and specificity, as well as the positive predictive value, negative predictive value, and area under the curve. Results: The rate of SSI was 9.33%. Using a backward stepwise approach, we identified that the remarkable risk factors predicting SSI in the multi-variate Cox regression analysis were age, body mass index, smoking, cerebrospinal fluid leakage, drain duration and pre-operative albumin level. Compared with other ML algorithms, the NB model had the highest performance in seven ML models, with an average area under the curve of 0.95, sensitivity of 0.78, specificity of 0.88, and accuracy of 0.87. Conclusions: The NB model in the ML algorithm had excellent calibration and accurately predicted the risk of SSI compared with the existing models, and might serve as an important tool for the early detection and treatment of SSI following spinal infection.