B-249 Leveraging Routine Laboratory Parameters for Pediatric Respiratory Infection Triage: A Two-Stage Interpretable Machine Learning Approach for Differentiating Mycoplasma pneumoniae and Influenza

Abstract

Pediatric respiratory tract infections (RTIs), primarily caused by influenza and Mycoplasma pneumoniae (MP), remain a global health priority. During peak seasons, the co-circulation of multiple pathogens makes clinical differentiation based on symptoms alone difficult. While routine laboratory tests—such as complete blood count (CBC) with integrated whole-blood C-reactive protein (fastCRP) analysis and basic biochemistry—are standard, their multidimensional diagnostic potential through machine learning (ML) has not been fully realized for rapid pathogen-specific triage. To develop and validate an interpretable two-stage ML framework that utilizes routine laboratory parameters to screen for broad respiratory infections and subsequently differentiate between influenza and MP in children. This retrospective study analyzed 8,504 clinical records from 2020 to 2024. After applying rigorous exclusion criteria—including individuals aged 18 years or older, recent COVID-19 infection, underlying immunodeficiencies, asthma, allergic diseases, immunosuppressive therapy, and incomplete laboratory data (missing CBC or biochemistry)—a refined cohort of 2,499 pediatric patients was utilized. The diagnostic workflow was structured as a two-stage cascade architecture. Stage 1 focused on distinguishing "negative" individuals from infection-positive patients. Negative cases were defined as those testing negative for 13 common respiratory pathogens: Influenza A (H1N1 and H3N2), Influenza B, Adenovirus, Bocavirus, Rhinovirus, Parainfluenza, Coronavirus, Respiratory Syncytial Virus (RSV), Metapneumovirus, MP, Chlamydia, and SARS-CoV-2. Patients identified as positive with a single-pathogen infection (specifically MP or Influenza) proceeded to Stage 2, which aimed to differentiate between these two most prevalent types. To address the significant class imbalance between the influenza group (n=133), MP group(n=475) and the negative control group (n=1,891), random undersampling was implemented during training. The modeling process utilized Random Forest (RF) for Stage 1 and LightGBM (LGBM) for Stage 2, selected after comparing performance against XGBoost and Logistic Regression. Feature selection involved a rigorous consensus approach: candidate predictors were identified through LASSO regression, univariate logistic regression, and feature importance analysis, with final features determined via Venn diagram intersection. Model hyperparameters were optimized using cross-validation on the training set. Performance was evaluated using AUC, accuracy, precision, recall, and F1-score, with clinical interpretability provided by SHAP (SHapley Additive exPlanations) analysis. In Stage 1, six key features—including fastCRP Calcium, Phosphorus, Absolute Lymphocyte Count, Anion Gap (AG), and Basophil percentage—were identified, with the RF model achieving a recall of 0.7955. For Stage 2, four features—Monocyte percentage, Magnesium, RDW and AG—were selected. The final integrated cascade model demonstrated robust performance on the independent test set (AUC 0.8317; Accuracy 0.7551; Sensitivity 0.7083; F1-Score 0.7542). SHAP analysis revealed fastCRP as the primary predictor for Stage 1 screening, while Anion Gap (AG) emerged as the most discriminative feature for Stage 2 differentiation. The proposed two-stage ML framework provides an interpretable and accessible tool for rapid pediatric triage. By integrating the significant predictive value of integrated whole-blood CRP for initial infection screening with routine metabolic markers like Anion Gap for MP and influenza differentiation, this study demonstrates a cost-effective, data-driven approach for enhancing clinical decision-making in laboratory medicine.