An interpretable machine learning model for predicting febrile seizures following enterovirus infection in children

Area of Science:

• Pediatric Infectious Diseases
• Computational Biology
• Machine Learning in Medicine

Background:

• Enterovirus (EV) infections are common in children.
• Febrile Seizures (FS) are a frequent complication of EV infections.
• Accurate risk prediction for FS is crucial for clinical management.

Purpose of the Study:

• To develop an interpretable machine learning model for predicting FS risk in children with EV infections.
• To implement the model for clinical application and risk stratification.
• To identify key clinical predictors of FS in this population.

Main Methods:

• Retrospective study of 446 children with EV infection.
• Feature selection using LASSO regression and BORUTA algorithm from 53 clinical variables.
• Development and evaluation of six machine learning models (XGBoost, logistic regression, KNN, Naive Bayes, MLP, random forest) using AUC, sensitivity, specificity, F1 score, and Decision Curve Analysis (DCA).
• Interpretability analysis using SHAP values.
• Development of a web-based calculator using a Shiny application.

Main Results:

• The XGBoost model achieved high predictive performance: training AUC 0.972, internal validation AUC 0.842.
• Key predictors identified include fever duration, disease course, immunoglobulin M, neutrophil count, fibrinogen, and various immune cell percentages.
• SHAP analysis confirmed model interpretability, and DCA supported its clinical applicability.
• A user-friendly web-based calculator was developed for personalized risk assessment.

Conclusions:

• The developed XGBoost model offers high accuracy and clinical interpretability for predicting FS risk in children with EV infections.
• The web-based calculator provides a valuable tool for risk stratification and guiding management decisions.
• This approach enhances the clinical utility of machine learning in pediatric infectious disease management.