Preventing Heterotopic Ossification in Combat Casualties-Which Models Are Best Suited for Clinical Use?

BACKGROUND: To prevent symptomatic heterotopic ossification (HO) and guide primary prophylaxis in patients with combat wounds, physicians require risk stratification methods that can be used early in the postinjury period. There are no validated models to help guide clinicians in the treatment for this common and potentially disabling condition. QUESTIONS/PURPOSES: We developed three prognostic models designed to estimate the likelihood of wound-specific HO formation and compared them using receiver operating characteristic (ROC) curve analysis and decision curve analysis (DCA) to determine (1) which model is most accurate; and (2) which technique is best suited for clinical use. METHODS: We obtained muscle biopsies from 87 combat wounds during the first débridement in the United States, all of which were evaluated radiographically for development of HO at a minimum of 2 months postinjury. The criterion for determining the presence of HO was the ability to see radiographic evidence of ectopic bone formation within the zone of injury. We then quantified relative gene expression from 190 wound healing, osteogenic, and vascular genes. Using these data, we developed an Artificial Neural Network, Random Forest, and a Least Absolute Shrinkage and Selection Operator (LASSO) Logistic Regression model designed to estimate the likelihood of eventual wound-specific HO formation. HO was defined as any HO visible on the plain film within the zone of injury. We compared the models accuracy using area under the ROC curve (area under the curve [AUC]) as well as DCA to determine which model, if any, was better suited for clinical use. In general, the AUC compares models based solely on accuracy, whereas DCA compares their clinical utility after weighing the consequences of under- or overtreatment of a particular disorder. RESULTS: Both the Artificial Neural Network and the LASSO logistic regression models were relatively accurate with AUCs of 0.78 (95% confidence interval [CI], 0.72-0.83) and 0.75 (95% CI, 0.71-0.78), respectively. The Random Forest model returned an AUC of only 0.53 (95% CI, 0.48-0.59), marginally better than chance alone. Using DCA, the Artificial Neural Network model demonstrated the highest net benefit over the broadest range of threshold probabilities, indicating that it is perhaps better suited for clinical use than the LASSO logistic regression model. Specifically, if only patients with greater than 25% risk of developing HO received prophylaxis, for every 100 patients, use of the Artificial Network Model would result in six fewer patients who unnecessarily receive prophylaxis compared with using the LASSO regression model while not missing any patients who might benefit from it. CONCLUSIONS: Our findings suggest that it is possible to risk-stratify combat wounds with regard to eventual HO formation early in the débridement process. Using these data, the Artificial Neural Network model may lead to better patient selection when compared with the LASSO logistic regression approach. Future prospective studies are necessary to validate these findings while focusing on symptomatic HO as the endpoint of interest. LEVEL OF EVIDENCE: Level III, prognostic study.

Lessons of War: Turning Data Into Decisions.

BACKGROUND: Recent conflicts in Afghanistan and Iraq produced a substantial number of critically wounded service-members. We collected biomarker and clinical information from 73 patients who sustained 116 life-threatening combat wounds, and sought to determine if the data could be used to predict the likelihood of wound failure. METHODS: From each patient, we collected clinical information, serum, wound effluent, and tissue prior to and at each surgical débridement. Inflammatory cytokines were quantified in both the serum and effluent, as were gene expression targets. The primary outcome was successful wound healing. Computer intensive methods were used to derive prognostic models that were internally validated using target shuffling and cross-validation methods. A second cohort of eighteen critically injured civilian patients was evaluated to determine if similar inflammatory responses were observed. FINDINGS: The best-performing models enhanced clinical observation with biomarker data from the serum and wound effluent, an indicator that systemic inflammatory conditions contribute to local wound failure. A Random Forest model containing ten variables demonstrated the highest accuracy (AUC 0.79). Decision Curve Analysis indicated that the use of this model would improve clinical outcomes and reduce unnecessary surgical procedures. Civilian trauma patients demonstrated similar inflammatory responses and an equivalent wound failure rate, indicating that the model may be generalizable to civilian settings. INTERPRETATION: Using advanced analytics, we successfully codified clinical and biomarker data from combat patients into a potentially generalizable decision support tool. Analysis of inflammatory data from critically ill patients with acute injury may inform decision-making to improve clinical outcomes and reduce healthcare costs. FUNDING: United States Department of Defense Health Programs.