Development and internal validation of a risk prediction model for ipsilateral upper-limb lymphedema following breast cancer surgery.

Background: Accurate prediction of breast cancer-related lymphedema (BCRL) is essential for identifying high-risk patients, guiding early preventive interventions, and improving postoperative quality of life among breast cancer survivors. This study aimed to develop and internally validate a clinically practical predictive nomogram for BCRL. Methods: This retrospective cohort study included 234 patients undergoing breast cancer surgery, of whom 27 (11.5%) developed BCRL. Candidate predictors were first screened using LASSO-based regression. Subsequently, clinically relevant variables were entered into multivariable analysis, and Firth penalized logistic regression was applied to reduce sparse-data bias and improve estimate stability in the setting of limited outcome events. A nomogram was then constructed based on the statistically significant predictors. Model performance was evaluated by receiver operating characteristic (ROC) analysis, calibration measures, bootstrap internal validation, and decision curve analysis (DCA). Results: The final nomogram incorporated four variables: surgery type, pectoral nodes dissection, number of harvested lymph nodes, and N stage. In the Firth penalized logistic regression model, breast-conserving surgery was associated with a lower risk of BCRL, whereas pectoral nodes dissection, higher N stage, and a greater number of harvested lymph nodes were associated with increased risk. The model demonstrated excellent discriminative ability, with an apparent area under the curve (AUC) of 0.964 and a bootstrap optimism-corrected AUC of 0.954. Calibration analysis showed a calibration-in-the-large of 0.000, an apparent calibration slope of 1.000, and an optimism-corrected calibration slope of 0.711, indicating residual overfitting. The apparent Brier score was 0.0469, with an optimism-corrected value of 0.0537. DCA showed that the nomogram provided favorable net benefit relative to the treat-all and treat-none strategies over much of the clinically relevant threshold range (5%-80%). Conclusion: We developed and internally validated a clinically practical nomogram for predicting BCRL using four readily available variables. The model showed strong discrimination, acceptable calibration, and potential clinical utility. Pending external validation, this tool should be regarded as a supportive risk-stratification aid rather than a stand-alone clinical decision-making instrument, and it may help identify patients who warrant closer surveillance, preventive counseling, or early rehabilitation referral.