An interpretable machine learning model for predicting febrile seizures following enterovirus infection in children.
A machine learning model accurately predicts febrile seizures in children with enterovirus infections, but further validation is needed before clinical implementation.
Where it sits
this study against the rest of the pt-141 (bremelanotide) corpusSummary and findings
This study developed a machine learning model to predict febrile seizures in children with enterovirus infections. The XGBoost model showed high predictive performance with an AUC of 0.972 in the training set. A web-based calculator was created for clinical application.
Abstract
<h4>Objective</h4>This study aims to develop an interpretable machine learning model for predicting the risk of FS in children with enterovirus (EV) infections and to implement it for clinical application.<h4>Methods</h4>This retrospective study included 446 hospitalized children with EV infection (144 FS, 302 non‑FS). LASSO regression and BORUTA algorithm selected 15 key predictors from 53 clinical variables. Six models (logistic regression, KNN, Naive Bayes, MLP, random forest, XGBoost) were constructed and evaluated using AUC, sensitivity, specificity, F1 score, and decision curve analysis. SHAP values provided interpretability, and a Shiny web‑based calculator was developed.<h4>Results</h4>The XGBoost model demonstrated the best predictive performance: the training set AUC reached 0.972 (95% CI: 0.958-0.987), with sensitivity of 0.892 and specificity of 0.905. The internal validation set achieved an AUC of 0.842 (95% CI: 0.757-0.926). DCA confirmed its strong clinical applicability. SHAP analysis identified key features contributing to the model: fever duration, disease course, immunoglobulin M, neutrophil count, fibrinogen, CD8+ T-cell percentage, aspartate aminotransferase, CD3+ T-cell percentage, procalcitonin, presence of hand-foot herpes lesions, CD19+ B-cell percentage, erythrocyte sedimentation rate, lymphocyte count, serum ferritin level, and alanine aminotransferase. The 'Shiny' calculator facilitates personalized risk assessment.<h4>Conclusion</h4>The XGBoost predictive model developed in this study demonstrated both high accuracy and clinical interpretability. The associated web-based calculator offers a new tool for risk stratification and management of FS in children with EV infections.
Background
Febrile seizures (FS) are a common complication in children with enterovirus (EV) infections, posing a significant clinical challenge. Accurate prediction of FS risk could improve management and outcomes. This study addresses the need for a reliable predictive model by leveraging machine learning techniques to identify key clinical variables associated with FS risk.
Methods
This retrospective study analyzed data from 446 hospitalized children with EV infections, of whom 144 experienced FS. The study employed LASSO regression and the BORUTA algorithm to select 15 key predictors from 53 clinical variables. Six machine learning models were constructed and evaluated, including logistic regression, KNN, Naive Bayes, MLP, random forest, and XGBoost. Model performance was assessed using AUC, sensitivity, specificity, F1 score, and decision curve analysis. SHAP values were used for model interpretability, and a Shiny web-based calculator was developed for clinical use.
Results
The XGBoost model demonstrated superior predictive performance with a training set AUC of 0.972 (95% CI: 0.958-0.987), sensitivity of 0.892, and specificity of 0.905. In the internal validation set, the AUC was 0.842 (95% CI: 0.757-0.926). Decision curve analysis supported the model's clinical applicability. SHAP analysis identified key predictors such as fever duration, disease course, and various immunological and biochemical markers.
Interpretation
The XGBoost model's high AUC indicates strong predictive capability, though the clinical significance of these findings requires further validation. The study's retrospective nature and single-cohort design limit generalizability. While the model offers a promising tool for FS risk stratification, external validation in diverse populations is necessary to confirm its utility in clinical practice.
Key findings
- 446 hospitalized children with EV infection were included.
- XGBoost model training set AUC was 0.972 (95% CI: 0.958-0.987).
- Internal validation set AUC was 0.842 (95% CI: 0.757-0.926).
- Sensitivity was 0.892 and specificity was 0.905.
- 15 key predictors were identified using SHAP values.
Limitations
- Retrospective study design
- Single cohort of hospitalized children
- No external validation
- Potential overfitting in model development
- Limited generalizability