Development of a Prognostic Naive Bayesian Classifier for Successful Treatment of Nonunions

Background: Predictive models permitting individualized prognostication for patients with fracture nonunion are lacking. The objective of this study was to train, test, and cross-validate a Bayesian classifier for predicting fracture-nonunion healing in a population treated with extracorporeal shock wave therapy. Methods: Prospectively collected data from 349 patients with delayed fracture union or a nonunion were utilized to develop a naive Bayesian belief network model to estimate site-specific fracture-nonunion healing in patients treated with extracorporeal shock wave therapy. Receiver operating characteristic curve analysis and tenfold cross-validation of the model were used to determine the clinical utility of the approach. Results: Predictors of fracture-healing at six months following shock wave treatment were the time between the fracture and the first shock wave treatment, the time between the fracture and the surgery, intramedullary stabilization, the number of bone-grafting procedures, the number of extracorporeal shock wave therapy treatments, work-related injury, and the bone involved (p < 0.05 for all comparisons). These variables were all included in the naive Bayesian belief network model. Conclusions: A clinically relevant Bayesian classifier was developed to predict the outcome after extracorporeal shock wave therapy for fracture nonunions. The time to treatment and the anatomic site of the fracture nonunion significantly impacted healing outcomes. Although this study population was restricted to patients treated with shock wave therapy, Bayesian-derived predictive models may be developed for application to other fracture populations at risk for nonunion.