Development and Validation of an Interpretable Machine Learning Model for Predicting Thrombocythemia Risk During...

Third-generation cephalosporins are widely used for severe infections but carry thrombocythemia risks complicating therapeutic decisions. Current predictive tools lack accuracy and clinical interpretability. This study aimed to develop an interpretable machine learning (ML) model for thrombocythemia risk stratification during cephalosporin therapy.

A retrospective cohort of 45,779 adults treated with third-generation cephalosporins (2019-2023) was analyzed. After exclusions (age <18, missing data, baseline platelet anomalies), 25,707 patients were included. Thrombocythemia was defined as platelet count >400×10⁹/L within 30 days post-treatment. Predictors encompassed demographics, comorbidities, medications, and laboratory parameters. Data preprocessing included multiple imputation and stratified partitioning (70% training, 30% testing). Three ML algorithms (XGBoost, Random Forest, LightGBM) were evaluated using ROC-AUC, Brier score, and clinical utility metrics. SHAP analysis provided model interpretability.

XGBoost demonstrated superior performance, achieving the highest test-set discrimination (AUC=0.858, 95% CI:0.814-0.902) and calibration (Brier score=0.0088). SHAP analysis identified Baseline platelet count (PLT), red blood cell count (RBC), creatinine (CRE), daily usage frequency, and sex as key drivers. PLT was the strongest predictor (SHAP range: -1.67 to +1.48), with lower PLT exerting protective effects. RBC and CRE ranked second and third in importance, showing nonlinear risk relationships. Key clinical interactions included amplified risk from malignancies (SHAP=-0.215) and protective effects of female sex (SHAP=-0.194).

This interpretable ML framework enables precise thrombocythemia risk prediction during cephalosporin therapy, balancing algorithmic performance with clinical actionability.