Development of a machine learning-based assistive diagnostic model for Epstein‑Barr virus‑associated hemophagocytic lymphohistiocytosis in pediatric patients.

BACKGROUND: Children with evolving Epstein‑Barr virus‑associated hemophagocytic lymphohistiocytosis (EBV-HLH) often do not fully meet the established HLH‑2004 diagnostic criteria at admission, making timely identification of progressors challenging and increasing the risk of poor outcomes. We aimed to develop an internally evaluated, interpretable assistive diagnostic model using routine admission parameters to differentiate EBV-HLH from uncomplicated EBV infection early. The model was compared against multiple machine learning classifiers, interpreted via SHapley Additive exPlanations (SHAP) analysis, and deployed as a web‑based research prototype to support clinical reasoning before full diagnostic criteria are formally met.

METHODS: A retrospective case-control study collected clinical and laboratory data from 244 pediatric inpatients with active EBV infection (Xi'an Children's Hospital, 2021-2025). HLH was diagnosed according to the HLH-2004 criteria (fulfilling ≥5 of 8 criteria). This study employed a time-matched case-control design, in which children with EBV-HLH were matched 1:3 with children with uncomplicated EBV infection based on the diagnosis date ±3 calendar days, then randomly split into training (171 cases) and validation (73 cases) sets at a 7:3 ratio. Significant univariate variables underwent least absolute shrinkage and selection operator (LASSO) regression analysis and Boruta feature selection. The final set of variables was analyzed using six machine learning classifiers-logistic regression (LR), support vector machine (SVM), random forest (RF), naive Bayes (NB), gradient boosting machine (GBM), and LASSO-with hyperparameters optimized for each. The best model was visualized via SHAP and deployed as a Shiny web calculator for individual use.

RESULTS: The study population had a median age of 6.4 years (interquartile range: 4.8-9.6) and 48.3% were male. Triglyceride (TG), interferon-γ (IFN-γ), serum ferritin (FE), platelet count (PLT), and the CD4+/CD8+ ratio are the five statistically significant features that have been validated and were retained for subsequent modeling. The RF model demonstrated excellent discriminative performance, yielding an area under the curve (AUC) of 0.895 [95% confidence interval (CI): 0.824-0.966] on the validation set, with a sensitivity of 0.735 and a specificity of 0.846. The SHAP ranked these features in descending importance: TG levels, IFN-γ levels, FE levels, PLT levels, and the CD4+/CD8+ ratio. A web calculator based on SHAP was deployed for individualized risk estimation for clinicians.

CONCLUSIONS: This work identified five key biomarkers for early differential identification of EBV-HLH in pediatric patients. The RF model showed promising internal discriminative ability, and SHAP confirmed its interpretability via individual‑level feature visualization. Due to the limitations of a single-center retrospective study design and the lack of external validation, this model is a preliminary exploratory tool and requires further validation in multicenter prospective cohorts before clinical translation.