Predicting Skeletal-Related Events in Bone Metastasis from Advanced Breast Cancer Using Machine Learning

The purpose of this study is to develop a machine learning (ML) model to predict skeletal-related events (SREs) in patients with bone metastasis from breast cancer.

Publicly available, patient-level data of patients with bone metastasis from breast cancer receiving zoledronic acid from a previous clinical trial was analyzed. Five feature sets (FS) and seven algorithms were utilized to develop ML models to predict SREs within 18 months. The model was trained with ten-fold cross-validation, repeated three times, and was evaluated through four-fold external cross-validation. Model performances were assessed by multiple metrics and the ability to differentiate cumulative risks of SREs. The model was interpreted by Shapley Additive Explanation.

Four hundred sixty cases with bone metastatic breast cancer were incorporated for ML model development. The ML model that utilized six features selected by the Boruta method and random forest algorithm demonstrated the numerically highest performance. (F1 score of 0.70) The mean absolute SHAP values suggested performance status, history of SREs, and serum alkaline phosphatase were the most important features. The ML model differentiated the cases with a high risk and a low risk of SREs, with median time for the first SRE of 248 days and 867 days, respectively. (Hazard ratio: 2.43 and 95% confidence interval: 1.86-3.18) CONCLUSION: A machine learning model to predict SREs in patients with bone metastatic breast cancer demonstrated the features related to SRE risk and its ability to identify the population with a high risk of SREs.