Predicting Escalation of Care for Childhood Pneumonia Using Machine Learning: Retrospective Analysis and Model...

Area of Science:

• Pediatric medicine
• Medical informatics
• Machine learning applications in healthcare

Background:

• Pneumonia is a major cause of mortality in children under five.
• Existing machine learning (ML) applications in pneumonia diagnostics have not focused on predicting care escalation in pediatric cases.
• This study addresses the need for ML-based clinical decision support for pediatric community-acquired pneumonia management.

Purpose of the Study:

• To develop a robust predictive tool for primary care physicians.
• To assist in determining optimal patient management and care setting.
• To predict the need for escalation of care in pediatric community-acquired pneumonia.

Main Methods:

• Retrospective analysis of 437 pediatric community-acquired pneumonia cases predating the COVID-19 pandemic.
• Encoding of clinical features from unstructured records using Integrated Management of Childhood Illness guidelines.
• Application of Synthetic Minority Oversampling Technique-Tomek for imbalanced data, Shapley additive explanations for feature selection, and hyperparameter tuning with ensembling for model optimization.

Main Results:

• Optimized models achieved 77%–88% accuracy in predicting the need for transfer to higher care levels.
• Area under the receiver operator characteristic curve (AUC-ROC) was 0.88, and area under the precision-recall curve (AUC-PR) was 0.96.
• Key predictors identified included hypoxia, respiratory distress, age, weight-for-age z score, and complaint duration, independent of lab diagnostics.

Conclusions:

• Machine learning techniques are feasible for creating prognostic tools in childhood pneumonia.
• The developed tool enables early identification of cases requiring escalated care.
• This approach combines clinical expertise with data science to enhance pediatric pneumonia management.