Clinical prediction models for pediatric Epstein-Barr virus infectious mononucleosis: from 6 machine learning algorithms Running Title: Machine learning Models for Pediatric EBV Infectious Mononucleosis

Abstract Background Pediatric Epstein-Barr virus (EBV) Infectious mononucleosis (PEBV-IM) is an acute infectious disease. However, there are no effective clinical diagnostic indicators for PEBV-IM. The study aimed to construct a clinical model for effective prediction of PEBV-IM and to identify relevant key feature variables, thereby providing a favorable clinical decision-making tool for PEBV-IM. Methods Data were obtained from the clinical diagnosis of PEBV-IM patients, and the feature variables were acquired by least absolute shrinkage and selection operator (LASSO) regression analysis. Subsequently, optimal clinical prediction models and key feature variables for PEBV-IM were acquired using 6 machine learning (ML) algorithms. Finally, to further investigate the relationship between optimal clinical prediction models and key feature variables, the SHapley Additive exPlanations (SHAP) model interpretation was proceeded. Results A total of 60 PEBV-IM samples and 41 variables were included in the analyses, and 12 characteristic feature variables were identified by LASSO. Subsequently, founded on the feature variables, the clinical prediction model was constructed using the plsRglm algorithm, which achieved the highest accuracy in both the training set (area under the curve (AUC) = 0.939) and the validation set (AUC = 0.850). Thus, this model was identified as the optimal clinical prediction model, while key feature variables platelet count and gamma-glutamyl transferase (GGT) were acquired. Notably, the GGT had the significant effect on the output of the clinical prediction model, with low GGT having a positive effect on the output, while low feature values of platelet count had a negative effect on the model output. Conclusion Obtaining a highly accurate clinical prediction model for PEBV-IM and 2 key feature variables (platelet count and GGT), which, in combination with SHAP model interpretation, provided a clear understanding and a novel tool for early diagnosis and clinical decision-making in PEBV-IM.