(877) A Supervised Machine Learning Model Predicting Pediatric Intensive Care Unit Mortality

Introduction: Accurate prediction of mortality in pediatric intensive care unit (PICU) patients is an important and challenging task for clinicians. In this study a supervised machine learning (ML) model was developed to predict PICU mortality.

Methods: Trichotomous Outcome Prediction in Critical Care (TOPICC) study dataset including patients 0 to 18 years of age admitted to general and cardiovascular PICUs between 2011 to 2013 was analyzed. The ML model used LightGBM (Light Gradient Boosting Machine) algorithm. The ML model’s input variables were the Pediatric Risk of Mortality (PRISM) III score components including (heart rate (HR), systolic blood pressure (SBP), temperature, pupillary reactivity, mental status, arterial PaO2, pH, PaCO2, bicarbonate, glucose, potassium, blood urea nitrogen (BUN), creatinine, white blood cell count, platelet count, prothrombin and partial thromboplastin time). These variables were collected within a 6-hour window around PICU admission. The ML model was trained on 80% of the dataset and tested on 20% of the dataset utilizing 10-fold cross validation. ML model’s mortality predictive performance was evaluated by accuracy, area under the receiver operating characteristic curve (AUC-ROC), precision (positive predictive value), recall (sensitivity), area under the precision-recall curve (AUC-PR), F1 score (harmonic mean of precision and recall) and Brier score. Input variable importance was evaluated by SHapley Additive exPlanations (SHAP) method.

Results: The dataset included 10,078 patients. The median age was 3.7 years (IQR: 0.8,10.8), and the mortality rate was 2.7%. The ML model achieved an accuracy of 0.98 (95% Confidence Interval (CI): 0.97, 0.98), AUC-ROC of 0.88 (95% CI: 0.82, 0.92), precision of 0.92 (95% CI: 0.70, 1.00), recall of 0.26 (95% CI: 0.12,0.38) AUC-PR of 0.40 (95% CI:0.26,0.57), F1 score of 0.4 (95% CI:0.22,0.56) and Brier score of 0.021 (95% CI: 0.015,0.029). SHAP analysis identified the lowest HR, lowest SBP, highest BUN, highest PaCO2 and highest body temperature as the top five most important predictors of mortality.

Conclusions: The supervised ML model showed promising potential for predicting PICU mortality. Future work should focus on enhancing recall and further validating the model across larger datasets to test its generalizability.