(655) Machine Learning Prediction of PICU-Associated Parental Posttraumatic Stress Symptoms

Introduction: Parental Post-Traumatic Stress Symptoms (PTSS) is a significant morbidity of a child’s admission to the Pediatric Intensive Care Unit (PICU). Prevalence rates range from 13-60% up to 30 months following a child’s PICU discharge. Despite known risk factors for parental PTSS we have limited inpatient methods for predicting risk of developing parental PTSS.

Methods: Using data of 169 subjects from a previous cohort. We considered Elastic Nets, lasso regression, random forests, XGBoost, and fully connected neural networks as predictive models for PICU associated parental PTSS. This XGBoost model had the best performance model on the test set as measured by the accuracy for binary classification across all possible thresholds. We computed the variable importance of all the predictions from the selected final model using a Local Interpretable Model-Agnostic Explanation (LIME) approach, identifying the 10 most important variables for each of the 169 data point prediction. We also summarized the direction of the variable's effect on the risk outcome by averaging and computing the sample standard deviation and the associated coefficient of the local linear regressions.

Results: The fitted model predicted subjects with PICU associated parental PTSS at 3-9 months with 76.7% accuracy, sensitivity and specificity of 0.83 (95% CI 0.58, 0.96) and 0.72 (95% CI 0.50, 0.87) respectively and precision of 0.68 (95% CI 0.45, 0.86). The area under the receiver operating curve was 0.78 (95% CI 0.64,0.92). The 3 most important variables associated with parental PTSS included prior traumatic experiences/mental health concerns (coefficient 0.19, 95% CI 0.19, 0.19), the prior plus parental concerns of patient death or permanent injury (coefficient 0.15, 95% CI 0.15,0.153) and diagnosis of Acute Stress Disorder at discharge (coefficient 0.11, 95% CI 0.10,0.11).

Conclusions: This model demonstrates major risk variables for PICU associated parental PTSS prediction. Using this model and human expertise to refine and validate an inpatient predictive tool will allow for investigation of various treatment options for PICU associated parental PTSS.