Development and external validation of a machine learning prediction model for Epstein-Barr virus-associated hemophagocytic lymphohistiocytosis in children using routine blood parameters: a retrospective cohort study.

内容摘要

Epstein-Barr virus (EBV) infection is a common pediatric infectious disease. Infectious mononucleosis (IM) and hemophagocytic lymphohistiocytosis (HLH), two major complications of EBV infection, share similar clinical manifestations in the early stage. While IM is typically self-limiting, HLH is life-threatening and requires immediate intervention. Early differentiation between these two conditions is crucial for clinical decision-making; however, reliable prediction models based on readily available laboratory parameters remain scarce. This study aimed to develop and validate a machine learning prediction model using routine blood parameters obtained within 24&#xa0;h of hospital admission in children with confirmed acute EBV infection to distinguish EBV-associated IM (EBV-IM) from EBV-associated HLH (EBV-HLH). This retrospective cohort study included 4,871 pediatric patients diagnosed with either EBV-IM or EBV-HLH from two campuses of Children's Hospital of Chongqing Medical University. Demographic information and initial complete blood count (CBC) parameters within 24&#xa0;h of admission were collected. The cohort was divided into a model development group (Yuzhong Campus, n&#x2009;=&#x2009;2,848; 70% for training, 30% for internal testing) and an external validation group (Liangjiang Campus, n&#x2009;=&#x2009;2,023). Thirteen machine learning algorithms were evaluated using random search with 5-fold cross-validation for hyperparameter tuning. Shapley Additive exPlanations (SHAP) analysis was performed to interpret model predictions. EBV-HLH accounted for 12.46% (607/4,871) of the total cohort, with significantly different prevalence between the development and validation cohorts (18.29% vs. 4.25%, p&#x2009;<&#x2009;0.001). Significant differences were observed between cohorts in age and all CBC parameters except gender (p&#x2009;<&#x2009;0.05). The Random Forest (RF) model demonstrated optimal performance in the internal validation set (AUC&#x2009;=&#x2009;0.993, 95% CI: 0.990-0.996). In the external validation cohort, the RF model maintained robust discriminative ability (AUC&#x2009;=&#x2009;0.971, 95% CI: 0.949-0.992). Calibration curves indicated excellent agreement between predicted probabilities and actual risks. SHAP analysis identified WBC, PLT, LAC, and Hb as the most critical predictors of EBV-HLH. DCA demonstrated substantial clinical net benefit. A free online decision-support tool (https://wangrj1988.shinyapps.io/EBV-HLH-IM/) was developed based on the RF model to facilitate real-time risk assessment. The RF-based model using routine CBC parameters enables admission-based risk assessment of pediatric EBV-HLH with excellent generalizability, offering a cost-effective tool for diverse healthcare settings. Clinical trial number: not applicable.