Explainable machine learning-based preliminary screening for viral encephalitis by blood routine analysis

Viral encephalitis (VE) is a severe neurological emergency; however, timely diagnosis remains challenging, particularly in resource-limited settings. This study aimed to develop and validate an interpretable machine learning (ML) model for the preliminary risk stratification of VE based solely on routine blood analysis (RBA).

A retrospective cohort of patients (n = 313, train/test = 8/2) with suspected VE was collected from the electronic health records of a tertiary hospital. ML models, including K-nearest neighbor (KNN), light-gradient boosting model (GBM), logistic regression, random forest, supporting vector machine (SVM), and extreme gradient boosting (XGBoost), were trained and validated using demographic and hematological features. The Shapley Additive Explanations (SHAP) framework was used to ensure model transparency and interpretability.

The XGBoost model demonstrated superior performance with an AUC of 0.949 (95%CI: 0.921 ~ 0.978) in 10-fold cross-validation in the train set and 0.900 (95% CI: 0.801-1.000) in the test set. SHAP analysis identified serum albumin (ALB) and white blood cell (WBC) counts, and low neutrophil (NEU) counts were the most significant contributors to VE prediction. Notably, the interactions between ALB and WBC were also highly influential in VE prediction.

This study presents an accurate and explainable XGBoost model for VE preliminary screening. Based on universally available blood indicators, it serves as a practical front-end tool for optimizing diagnostic workflows in emergency or primary care settings.