Integrating Serum Lipid Biomarkers Into Machine Learning for the Differential Diagnosis of Breast Nodules

Abstract

Objective To develop and validate an interpretable machine learning (ML) model for predicting malignant risk in patients with breast nodules using serum lipid biomarkers. Methods This retrospective study included 899 patients with breast nodules (236 malignant) admitted between March 2022 and December 2024. Patients were randomly assigned to a training cohort (n = 630) and an internal validation cohort (n = 269) at a 7:3 ratio. Baseline clinical and laboratory data were collected upon admission. Following feature selection via LASSO regression, the predictive performance of 8 ML algorithms was evaluated and compared using receiver operating characteristic (ROC) curves. The optimal model’s performance was further corroborated using an independent temporal validation cohort of 190 patients (admitted Jan–Aug 2025). Model interpretability was addressed using SHapley Additive exPlanations (SHAP). Results Nine key predictors were identified from 20 candidates by Lasso regression and clinical expertise. The random forest (RF) model outperformed other algorithms, achieving areas under the curve (AUC) values of 0.789, 0.782, and 0.825 for the training, internal validation, and temporal validation cohorts, respectively. Hosmer–Lemeshow tests (p > 0.05) indicated high calibration between the predicted and observed risks. SHAP importance analysis revealed Fer, age, and CEA to be the top three predictive factors. Conclusion The RF model based on serum lipid biomarkers serves as a robust, noninvasive tool for assessing breast cancer risk, showing significant potential for clinical decision support in screening programs.