Development and validation of an interpretable machine learning model integrating inflammatory blood biomarkers for predicting postoperative breast cancer-related lymphedema: A prospective cohort study.

BACKGROUND: Breast cancer-related lymphedema is a frequent and debilitating complication. Early identification of high-risk individuals is critical, yet conventional risk models rarely incorporate systemic inflammatory status. METHODS: In this prospective cohort, 880 patients undergoing surgery for primary breast cancer were enrolled. Preoperative inflammatory blood biomarkers and clinicopathologic data were collected. A machine learning-based inflammatory model was developed using least absolute shrinkage and selection operator for feature selection and six supervised algorithms for model construction. Independent clinical predictors were identified using multivariable logistic regression, and a combined model integrating inflammatory signature and clinical factors was constructed. Model performance was evaluated in training and test cohorts using discrimination, calibration, and decision curve analysis. Model interpretability was assessed using SHapley Additive exPlanations, and a nomogram was generated for individualized risk calculation. RESULTS: Overall, 29.0% of patients developed breast cancer-related lymphedema. Fourteen inflammatory biomarkers were selected to construct the signature, and the support vector machine-based inflammatory model demonstrated robust predictive performance (area under the receiver operating characteristic curve: 0.861 in the training set, 0.781 in the test set). Body mass index and lymphadenectomy were identified as independent clinical predictors. The combined model achieved improved discrimination (area under the curve: 0.871 in the training set, 0.811 in the test set), demonstrated favorable calibration and clinical utility. SHapley Additive exPlanations analysis identified platelet × fibrinogen, neutrophil × platelet, derived neutrophil-to-lymphocyte ratio, systemic inflammation response index, and fibrinogen-to-lymphocyte ratio as the most influential features. The nomogram enables individualized risk estimation to guide surveillance and early intervention. CONCLUSIONS: Integrating systemic inflammatory profiling with clinical variables enables accurate and interpretable prediction of breast cancer-related lymphedema, supporting personalized monitoring and early preventive strategies to reduce postoperative morbidity.