Learning Objective 1: Discuss methods and strategies for incorporating research results into predictive models.
Learning Objective 2: Describe innovative processes for translating research findings and predictive models into clinical practice.
Methods: Regression models and data mining identified predictors of procedural distress. Predictive models were built using support vector machine (SVM) regression and further improved by automatic feature selection. Each model was extensively evaluated using multiple cross-validation tests. These data were incorporated into a computerized decision support system in the next phase of study.
Results: In the initial study, measures of procedural distress for 542 children included child-report, parent-report, behavioral observation, and a biomarker. From 255 initial items, regression models identified 30 items explaining child distress, whereas data mining identified 44 predictive items. In multiple cross-validation tests, the predictive accuracy of data mining was superior to linear regression (p < 0.01). In the second phase of the trial, the model predicted level of distress for 582 children, and informed the level of intervention received. For each intervention group, there was a significant difference (p < 0.006) between dyads predicted to be low vs. high risk for distress, validating effective identification of risk assignment.
Conclusion: Findings from this multi-phase study resulted in the development of new tools to predict child procedural distress and ultimately inform clinical interventions to reduce distress. Current work includes re-training the predictive model using data from the second phase and refining the computer application for translation into clinical practice.