Implementing Symptom-Based Predictive Models for Early Diagnosis of Pediatric Respiratory Viral Infections

Area of Science:

• Pediatric Infectious Diseases
• Computational Epidemiology
• Machine Learning in Healthcare

Background:

• Common pediatric respiratory viral infections include SARS-CoV-2, RSV, influenza, rhinovirus, and adenovirus.
• Acute respiratory infections (ARIs) are a significant health concern in children.
• Symptom-based predictive models can expedite ARI diagnosis in primary care.

Purpose of the Study:

• To evaluate the efficacy of machine learning models in estimating infection probabilities for common pediatric respiratory viruses using symptom data.
• To assess the diagnostic accuracy of these models in a primary care setting.

Main Methods:

• Utilized data from 868 children with ARI symptoms across 14 primary care centers (COPEDICAT cohort).
• Employed random forest and boosting models with 10-fold cross-validation and SMOTE-NC for class imbalance.
• Evaluated model performance using AUC, sensitivity, specificity, and SHAP values for feature importance.

Main Results:

• Models demonstrated higher accuracy for Respiratory Syncytial Virus (RSV) (AUC: 0.81) and influenza viruses (AUC: 0.71).
• Effectively ruled out SARS-CoV-2 infections based on the absence of specific symptoms like crackles and wheezing.
• Lower predictive performance for rhinovirus and adenovirus due to nonspecific symptoms; SHAP analysis identified key symptom patterns for each virus.

Conclusions:

• Symptom-based predictive models are effective tools for identifying pediatric respiratory infections.
• Models show notable accuracy for RSV, SARS-CoV-2, and influenza virus infections.
• Machine learning aids in differential diagnosis of viral ARIs in children based on clinical presentation.