Keywords: Clinical Decision Support Tool, Machine Learning, Biomarkers, High Dimensional Data, Kaplan-Meier
Data analytics is at the leading edge of clinical care research. Novel data-driven predictive models aim to deliver precision medicine to patients in numerous health care settings. Our group at the Uniformed Services University of the Health Sciences Surgical Critical Care Initiative developed a clinical decision support tool (CDST) that visualizes complex health data to estimate successful and unsuccessful surgical closure of complex traumatic wounds. Here we investigated an approach to train effective predictive models that estimate successful and unsuccessful surgical closure of traumatic wounds.
We evaluated the treatment of 73 patients with 116 combat related extremity wounds, which involves multiple debridement operations before definitive closure of each wound. In each debridement, samples of serum and wound effluent were collected and tested for the level of 32 cytokines associated with wound healing. The method combined three statistical techniques: (1) the max-min parents and children algorithm to calculate the nodes of Bayesian Networks as a reduced variable set; (2) a random forest for classification of dehiscence and healing; and (3) the Kaplan-Meier curve for visualization of joint probabilities of closure and successful or unsuccessful wound healing. We trained random forest models only on debridements with both successful and unsuccessful wound healing. We computed a Kaplan-Meier curve for each debridement. The random forest predictions and the Kaplan-Meier curves were multiplied by each other to produce a scaled Kaplan-Meier curve of each wound’s present and future joint probabilities of closure and successful or unsuccessful wound healing.
Cross-validated models trained using reduced variable sets calculated on each debridement classifying both successful and unsuccessful wound closure had accuracies of 0.79 to 0.9, Kappas of 0.35-0.7, sensitivities of 0.42 to 0.89 and specificities of 0.9 to 0.96. The scaled Kaplan-Meier curve analysis produced a graphic of estimated outcomes. Decision curve analysis confirmed reduced models outperform, or perform similarly to, models trained with the full variable set.
Combined multivariate analyses may be an effective way to evaluate healing of combat-related extremity wounds. This work highlights a potential approach to CDST development for clinician-driven interactive tools. Prospective developments include other clinical outcomes, more data and external validation sets.