Machine learning identifies EBV-associated HLH from routine labs

For practitioners building diagnostic ML for acute pediatric infections, these studies demonstrate that models using routine admission labs can reach high discrimination and remain interpretable, shifting emphasis toward early-available biomarkers and explainability for clinical uptake.

BMC Medical Informatics and Decision Making (Yingying Ye et al., published 27 June 2026) developed multiple diagnostic models and reports that an XGBoost classifier trained on 1,026 children with confirmed acute EBV infection achieved AUC 0.9775, sensitivity 0.9461, and specificity 0.9784, with SHAP analysis ranking D-dimer, cervical lymphadenopathy, GGT, LDH, and CD3+CD4+ T cells as the most important features. BMC Infectious Diseases (Li Xiao et al., published 27 June 2026) presents a larger retrospective cohort from two campuses totaling 4,871 pediatric patients, reports EBV-HLH prevalence of 12.46%, evaluated 13 machine learning algorithms with 5-fold cross-validation and random search hyperparameter tuning, and performed external validation across campuses using routine complete blood count parameters obtained within 24 hours of admission.

Both reports rely on routinely collected clinical and laboratory data within 24 hours, prioritizing features available at first contact. The use of XGBoost and ensemble-tree methods aligns with prior clinical-ML work where tabular data and class imbalance are common. Both teams applied SHAP for post-hoc explanation; this choice supports per-patient interpretability and helps surface laboratory markers that drive risk scores. Neither paper claims causality; both present predictive discrimination and feature importance as evidence for early diagnostic support.

The Soochow University cohort (Ye et al.) used LASSO for feature selection before training six ML algorithms plus logistic regression, selecting XGBoost as top performer based on AUC, sensitivity, and specificity. The Chongqing cohort (Li Xiao et al.) split data by campus for development (Yuzhong campus, n=2,848, 70% train/30% internal test) and external validation (Liangjiang campus, n=2,023), evaluated 13 algorithms, and used SHAP for model interpretation. Both studies emphasize early labs; Chongqing focuses on complete blood count parameters, while Soochow includes additional immunologic and biochemical markers such as CD3+CD4+ counts and D-dimer.

Observed patterns in similar diagnostic-ML deployments show two recurring requirements: rigorous external validation across sites to avoid dataset bias, and interpretable explanations to aid clinician trust. These studies provide both elements at different scales: one emphasizes a broad, routine-lab approach with multi-campus external validation (Chongqing), and the other demonstrates very high discrimination with a richer feature set and explicit SHAP-ranked biomarkers (Soochow).

Models trained on early, routinely available labs can achieve high discrimination for rare but severe complications like EBV-HLH, which supports building triage-oriented pipelines that ingest first-pass CBC and basic biochemistry. SHAP-style explanations allow mapping model predictions to specific markers such as D-dimer and LDH, which can be surfaced in clinical decision support without revealing internal reasoning beyond the evidence.

Indicators for real-world utility include prospective evaluation, calibration across age groups and comorbidity strata, integration into electronic health records with real-time lab feeds, and assessment of false-positive workload for hematology services. Additionally, multi-center prospective validation would address spectrum bias and verify that SHAP-identified features generalize beyond the reported cohorts.

Both articles are retrospective and note the need for prospective testing; differences in available features across centers (routine CBC-only versus extended panels) limit direct replication without harmonized input sets. Neither study provides prospective impact analysis on clinical outcomes.