Development and validation of an interpretable machine learning-based predictive model for breast cancer bone metastasis.

Background: Breast cancer is one of the most common malignancies worldwide, with bone metastasis representing its most frequent distant metastatic form, significantly worsening patient prognosis. This study aims to develop a machine learning-based predictive model for accurately assessing the risk of bone metastasis in breast cancer patients, thereby enabling personalized risk stratification, early clinical intervention, and optimized treatment strategies. Methods: This study utilized the Surveillance, Epidemiology, and End Results database as the primary data source to develop machine learning models for predicting bone metastasis risk in breast cancer patients. Initially, univariate and multivariate logistic regression analyses were conducted to screen key predictive variables; subsequently, eight machine learning algorithms were constructed based on the screening results.10-fold cross-validation employed for hyperparameter optimization. Following training, model performance was evaluated on an internal test cohort and externally validated on 342 real-world cases from an independent hospital cohort. Model assessment incorporated multiple metrics, including area under the curve (AUC), area under the precision-recall curve (AUPRC), decision curve analysis, and calibration curves. Additionally, SHAP analysis was applied to enhance model interpretability, and a web-based calculator was developed based on the optimal model to facilitate clinical application and decision support. Results: Baseline characteristics across cohorts indicated that the majority of patients were aged over 50 years, female, predominantly with the HR+/HER2- molecular subtype, and exhibited a low incidence of bone metastasis. Univariate and multivariate logistic regression analyses identified key independent risk factors, including age >50 years, higher tumor grade, advanced T stage, N stage, clinical stage and HR-/HER2- subtype factors included radiotherapy, surgery, and married status. The LGB model demonstrated superior performance, achieving an AUC of 0.98 in the training set and 10-fold cross-validation (standard deviation=0.00), 0.98 in the internal validation set, and 0.91 in the external validation set; AUPRC values across the three cohorts were 0.96, 0.79, and 0.87, respectively; decision curve analysis showed excellent net clinical benefit within the 0.1-0.8 threshold range; calibration curves further confirmed high concordance between predicted probabilities and actual event rates. SHAP analysis highlighted surgery as the primary protective factor, followed by N stage, T stage, and radiotherapy as risk enhancers; for example, advanced N stage was associated with positive SHAP values, indicating a significant increase in bone metastasis risk. Conclusions: This study developed an interpretable LGB model accompanied by a web-based calculator, thereby advancing personalized risk stratification, early detection of bone metastasis and optimized treatment strategies.