Machine learning identifies EBV-associated HLH from routine labs

BMC Medical Informatics and Decision Making (Ye et al., Children's Hospital of Soochow University, published June 27, 2026) trained six machine learning algorithms plus logistic regression on 1,026 children hospitalized with confirmed EBV infection between October 2017 and September 2024. The XGBoost model performed best, reaching AUC 0.9775, sensitivity 0.9461, and specificity 0.9784. SHAP analysis ranked D-dimer, cervical lymphadenopathy, GGT, LDH, and CD3+CD4+ T-cell count as the top predictors. Separately, BMC Infectious Diseases (Xiao et al., Children's Hospital of Chongqing Medical University, published June 27, 2026) evaluated 13 algorithms on a larger, two-campus cohort of 4,871 pediatric patients (development cohort n=2,848; external validation cohort n=2,023 from a separate campus). EBV-HLH prevalence was 12.46% overall but varied sharply between cohorts (18.29% development vs. 4.25% validation, p<0.001). A Random Forest model reached AUC 0.993 (95% CI 0.990-0.996) internally and AUC 0.971 (95% CI 0.949-0.992) on external validation, using only four routine complete blood count parameters: white blood cell count, platelet count, lymphocyte absolute count, and hemoglobin. The team released a free online calculator implementing the model.

Both teams used tree-based ensemble methods well suited to tabular clinical data and class imbalance, and both applied SHAP for per-patient interpretability rather than relying on black-box scores. The Soochow model used LASSO regression for feature selection before comparing six algorithms; the Chongqing model tuned 13 algorithms via 5-fold cross-validation with random search and validated the winning Random Forest model on an entirely separate hospital campus, the more demanding form of external validation. The Chongqing model needed only four routine CBC values, while the Soochow model drew on a richer panel including immunologic markers like CD3+CD4+ counts - a tradeoff between minimal-input deployability and predictor richness that clinical-ML teams building similar triage tools will need to weigh.

Diagnostic ML for rare, high-stakes pediatric complications faces a recurring bar: external validation across independent sites to rule out single-center overfitting, and interpretable outputs clinicians can act on quickly. The Chongqing study clears the higher bar of true external validation (a geographically distinct campus with a very different HLH prevalence), while the Soochow study demonstrates a stronger single-site AUC with a broader lab panel. Together they represent two credible but different paths toward the same clinical goal.

Both papers show that first-pass admission data, whether a narrow four-parameter CBC panel or a broader immunologic/biochemical panel, can support high-discrimination triage before specialist labs return. Teams building similar pipelines should note the prevalence shift between the Chongqing cohorts (18.29% vs. 4.25%) as a reminder that calibration, not just discrimination, needs to be checked across deployment sites with different base rates.

Both studies are retrospective; neither reports prospective clinical-outcome validation. Watch for prospective deployment of the Chongqing team's public risk calculator, further external validation beyond the two reported campuses, and whether either model's biomarker set generalizes to non-Chinese pediatric populations, since both cohorts are drawn from a single national health system.