Mortality risk model to predict bacterial infection in the paediatric intensive care unit

Abstract

Abstract Background Bacterial infections have historically posed a substantial threat to the survival of children in intensive care unit. Predicting high mortality risk in children due to bacterial infections is crucial for prevention and management, but there is currently no effective predictive method. In this study, we investigated a novel approach to address this challenge.Methods We utilized the Paediatric Intensive Care (PIC) database for this study. Hospitalised children with positive bacterial cultures were divided into three groups: positive culture, gram-positive, and gram-negative groups. We extracted data on demographics, vital signs, and laboratory analyses within 24 h of admission. The least absolute shrinkage and selection operator (LASSO) regression and XGBoost algorithms were employed to select and rank important features, and a logistic regression (LR) algorithm was used for model development with varying numbers of features. Evaluation of the models was done using receiver operating characteristic (ROC) curve analysis and compared against the Paediatric Mortality Risk Score III (PRISM III), Paediatric Logistic Organ Dysfunction Score-2 (PELOD-2), and Paediatric Multiple Organ Dysfunction Score (P-MODS).Results A total of 3695 children with bacterial infections were included. We developed LR models for three distinct groups of infection separately. These models demonstrated superior performance in predicting mortality risk compared to the PRISM III, PELOD-2, and P-MODS, achieving ROC-AUC(Receiver Operating Characteristic - Area Under the Curve) scores over 0.70.Conclusion This study created models for forecasting mortality risk in children with bacterial infections. The final models outperform traditional scoring models in predicting mortality while utilising fewer features. Additionally, these models are more comprehensible for paediatricians.