Machine Learning-Based Early Prediction of Pediatric Clinical Deterioration Using Multivariate Physiological and Laboratory Data

Abstract

Clinical deterioration in children can be a rapid process that necessitates rapid recognition to avoid adverse outcomes including respiratory failure, hemodynamic instability, need for transfer to intensive care, cardiac arrest, or death. This study aimed to create and test machine learning models for early prediction of pediatric clinical deterioration, based on multivariate physiological, laboratory and short-term trends. For analysis, a patient population of 800 paediatric patient records was created, with 171 (21.4%) patients having deteriorated within 24 hours. Demographic parameters, HR, RR, oxygen sat, SBP, temperature, CFT, GCS, urine output, inflammatory and metabolic lab parameters, and 6-hour trends were used as predictor variables. The five supervised learning algorithms logistic regression, decision tree, random forest, support vector machine and gradient boosting/XGBoost – were tested with accuracy, sensitivity, specificity, precision, F1 score, and area under the receiver operating characteristic curve. There was a correlation between deterioration and deteriorating respiratory, cardiovascular, neurological, perfusion, inflammatory and metabolic parameters. Logistic regression was the best-performing combination of sensitivity, F1, and interpretability, while gradient boosting was best at overall discrimination. The results illustrate the feasibility of combining physiological, lab, and temporal data to develop an early pediatric risk stratification. To enable clinical use, however, validation with real-world multicenter pediatric data is needed. These models help to escalate monitoring and treatment earlier.