An explainable machine learning model predicts pediatric varicella encephalitis.

BACKGROUND

Pediatric varicella encephalitis is a rare but serious complication of varicella, which has a significant impact on patient prognosis. Early clinical diagnosis is still challenging due to atypical clinical symptoms and lack of specific biomarkers. This study aims to establish a predictive model for pediatric varicella encephalitis and provide a practical tool for early clinical identification of such patients.

METHODS

A retrospective analysis method was used in this study. A total of 201 children with varicella were enrolled, including 156 in the training group and 45 in the testing group. LASSO regression, XGBoost and random forest algorithm were used to screen key features, and prediction models were constructed based on 6 algorithms. The discrimination, calibration and clinical applicability of the models were verified by the testing set. Shapley additive interpretation (SHAP) analysis was used to interpret the models.

RESULTS

Six characteristic variables associated with pediatric varicella encephalitis were screened out, among which the random forest model showed excellent predictive performance with an area under the curve of 0.950 (95% confidence interval: 0.948-0.952). The calibration curve confirmed that the model was well calibrated, and decision curve analysis showed that it had high clinical utility and provided the greatest net benefit within the risk threshold range. SHAP analysis showed that rash duration, headache, and vomiting were the main characteristics affecting the occurrence of varicella encephalitis in children. In addition, the study created a clinical web application for real-time risk stratification of patients and personalized risk contributions visualized through SHAP.

CONCLUSION

This study identified 6 important clinical variables of pediatric varicella encephalitis, and the constructed random forest model can accurately and rapidly identify children with varicella encephalitis, which has important clinical application value for early clinical intervention.