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

Area of Science:

• Neurology
• Medical Informatics
• Biochemistry

Background:

• Viral encephalitis (VE) poses a significant neurological emergency.
• Timely diagnosis of VE is challenging, especially in resource-limited settings.
• Routine blood analysis (RBA) offers a potential avenue for preliminary risk stratification.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) model for the preliminary risk stratification of VE.
• To utilize routine blood analysis (RBA) data exclusively for VE risk assessment.
• To enhance diagnostic workflows in emergency and primary care settings.

Main Methods:

• A retrospective cohort of 313 patients with suspected VE was analyzed.
• Multiple ML models (KNN, GBM, logistic regression, random forest, SVM, XGBoost) were trained and validated.
• The Shapley Additive Explanations (SHAP) framework ensured model interpretability.

Main Results:

• The XGBoost model achieved high performance (AUC train: 0.949, AUC test: 0.900).
• Key predictors identified by SHAP analysis included serum albumin (ALB), white blood cell (WBC) counts, and low neutrophil (NEU) counts.
• Interactions between ALB and WBC significantly influenced VE prediction.

Conclusions:

• An accurate and explainable XGBoost model for preliminary VE screening was developed.
• The model leverages universally available blood indicators for practical application.
• This ML tool can optimize diagnostic workflows in emergency and primary care settings.