494 - Routine Laboratory Blood Tests Predict In-ICU Mortality in Pediatric Sepsis Patients: A Retrospective Analysis of the PIC Database

Publication Number: 494.6416

Hang Xing, Women & Infants Hospital of Rhode Island, Providence, RI, United States; Junfeng Li, luohe central hospital, luohe, Henan, China (People's Republic); Ting Wang, Henan Cancer Hospital, Zhengzhou, RI, United States; Xiaodi F. Chen, The Warren Alpert Medical School of Brown University, Barrington, RI, United States

Background: Sepsis is a life-threatening condition with high mortality rates, especially among critically ill pediatric patients in intensive care settings.

Objective: To develop a prediction model of in-ICU mortality in pediatric sepsis patients using routine blood tests that are easily accessible upon ICU admission.

Design/Methods: We conducted a retrospective analysis of pediatric sepsis patients from the pediatric-specific intensive care (PIC) database, focusing on patients recorded between 2010 and 2018 at the Children's Hospital of Zhejiang University School of Medicine. Lasso regression was applied to identify the most significant features associated with in-ICU mortality. Subsequently, machine learning algorithms, including Logistic Regression (LR), XGBoost (XGB), Random Forest (RF), and Deep Neural Network (DNN), were used to build prediction models that incorporated the selected features along with age and gender. Model performance was evaluated using Receiver Operating Characteristic (ROC) curves, while Decision Curve Analysis (DCA) assessed the clinical utility of each model across various threshold probabilities.

Results: The study cohort included 13,449 pediatric patients at the Children's Hospital of Zhejiang University School of Medicine, of whom 299 met the criteria for sepsis, with 35 patients experiencing in-ICU mortality. Lasso regression identified several routine blood test parameters as significant predictors of in-ICU mortality, including neutrophil percentage, platelet distribution width, monocyte count, lymphocyte percentage, hematocrit, plateletcrit, mean corpuscular volume (MCV), eosinophil count, white blood cell (WBC) count, and red cell distribution width (RDW) (P < 0.05). Among the predictive models, the Random Forest model demonstrated the highest predictive accuracy for in-ICU mortality (AUC = 0.7565), followed by XGBoost (AUC = 0.7497) and Logistic Regression (AUC = 0.7259). DCA curves indicated that XGBoost provided a higher net benefit than other models within the 7%-30% threshold probability range.

Conclusion(s): These findings suggest that combining routine blood tests with age and gender yields an accessible and reliable set of predictive features for in-ICU mortality in pediatric sepsis patients. Prioritizing sepsis patients identified as high-risk based on these features for closer monitoring and timely intervention may help reduce mortality rates in this vulnerable population.

(A) Classification performance of Logistic Regression (LR), XGBoost (XGB), Random Forest (RF), and Deep Neural Network (DNN) for predicting in-ICU mortality. (B) Decision Curve Analysis (DCA) illustrating the clinical utility of the prediction models (LR, XGB, RF, and DNN) across a range of threshold probabilities.
figure.pdf