A Dual-Model Machine Learning Approach Using Routine Clinical, Laboratory, and Cerebrospinal Fluid Parameters for Early Differentiation of Autoimmune from Viral Encephalitis

Abstract

Diagnostic tests for autoimmune encephalitis (AE) and viral encephalitis (VE) are often unavailable in resource-limited settings, delaying life-saving treatment decisions—a gap relevant to Sustainable Development Goal 3, Good Health and Well-being. This study provides a machine learning benchmark for early AE-versus-VE differentiation using routinely available data. We reanalyzed a public dataset of 75 patients (37 AE, 38 VE), using only routine clinical, peripheral blood, and cerebrospinal fluid (CSF) variables and deliberately excluding specialized immunophenotyping markers. LASSO logistic regression and a depth-constrained random forest were developed within a repeated stratified 5-fold cross-validation framework, with interpretability assessed via LASSO odds ratios, feature importance, and SHAP values. Fever, headache, and elevated lumbar puncture opening pressure significantly favored VE, while psychiatric symptoms and cognitive impairment favored AE (all p < 0.05). The random forest achieved a mean cross-validated ROC-AUC of 0.730 and threshold accuracy of 0.640, modestly outperforming LASSO (AUC = 0.655; accuracy = 0.520). Interpretability methods converged: fever, headache, elevated CSF white blood cell count, and LP opening pressure favored VE, while elevated CSF protein, CSF glucose, psychiatric symptoms, and lymphocyte percentage favored AE. Machine-learning models achieved modest-to-moderate discrimination using only routine variables; external validation and calibration are needed before clinical use.