Development of a prediction model for infectious mononucleosis using machine learning algorithms based on blood cell...

Infectious mononucleosis (IM) presents with nonspecific clinical manifestations, leading to frequent misdiagnosis or delayed diagnosis, and is associated with potentially severe complications. Existing etiological diagnostic methods are characterized by prolonged turnaround times. This study aims to establish an IM model using eight machine learning algorithms and select the optimal one, so as to further improve the laboratory diagnostic accuracy of IM.

The study included 234 patients diagnosed with IM from April 2024 to December 2025 as the case group. The control group comprised 478 non-IM subjects, consisting of 236 patients with other pathogenic infections who exhibited reactive lymphocytes on microscopic examination and 242 healthy individuals. Recursive feature elimination (RFE) in conjunction with cross-validation was employed to rank feature importance and select optimal feature variables. Eight machine learning algorithms were trained, and their predictive performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, Precision-Recall (PR) curve and Confusion Matrix. The contribution of each feature to the model's predictions was quantified using SHapley Additive exPlanation (SHAP) analysis. Subsequently, the model's performance was rigorously externally validated using an independent validation cohort.

Three features - Reactive lymphocyte percentage (Reactive lymph%), Lym-Y, and platelet-to-lymphocyte ratio (PLR) - were selected for constructing the IM predictive model. The model constructed using the Adaptive Boosting Classifier(AdaBoost) machine learning algorithm demonstrated the best performance in the test set. It achieved an AUC of 0.928, an accuracy of 0.853, a sensitivity of 0.898, a specificity of 0.832, and an F1 score of 0.800. SHAP consistent with decision feature importance rankings, indicated that Reactive lymph% was the most significant feature in the predictive model; It was associated with an elevated risk of IM. Validation cohort also confirmed the robust performance of the AdaBoost predictive model, with an AUC of 0.923 and an F1 score of 0.797.

The IM predictive model constructed based on the AdaBoost machine learning algorithm combined with three blood cell analysis parameters exhibits satisfactory predictive efficacy. Based on the established model, the missed laboratory diagnosis rate in IM may potentially be reduced, pending prospective validation.