Frontiers in Oncology (Aug 2022)

Osteoporosis, fracture and survival: Application of machine learning in breast cancer prediction models

  • Lichen Ji,
  • Lichen Ji,
  • Lichen Ji,
  • Lichen Ji,
  • Wei Zhang,
  • Wei Zhang,
  • Wei Zhang,
  • Xugang Zhong,
  • Xugang Zhong,
  • Xugang Zhong,
  • Tingxiao Zhao,
  • Tingxiao Zhao,
  • Tingxiao Zhao,
  • Xixi Sun,
  • Senbo Zhu,
  • Senbo Zhu,
  • Senbo Zhu,
  • Senbo Zhu,
  • Yu Tong,
  • Yu Tong,
  • Yu Tong,
  • Junchao Luo,
  • Junchao Luo,
  • Junchao Luo,
  • Junchao Luo,
  • Youjia Xu,
  • Di Yang,
  • Di Yang,
  • Di Yang,
  • Yao Kang,
  • Yao Kang,
  • Yao Kang,
  • Jin Wang,
  • Jin Wang,
  • Jin Wang,
  • Qing Bi,
  • Qing Bi,
  • Qing Bi

DOI
https://doi.org/10.3389/fonc.2022.973307
Journal volume & issue
Vol. 12

Abstract

Read online

The risk of osteoporosis in breast cancer patients is higher than that in healthy populations. The fracture and death rates increase after patients are diagnosed with osteoporosis. We aimed to develop machine learning-based models to predict the risk of osteoporosis as well as the relative fracture occurrence and prognosis. We selected 749 breast cancer patients from two independent Chinese centers and applied six different methods of machine learning to develop osteoporosis, fracture and survival risk assessment models. The performance of the models was compared with that of current models, such as FRAX, OSTA and TNM, by applying ROC, DCA curve analysis, and the calculation of accuracy and sensitivity in both internal and independent external cohorts. Three models were developed. The XGB model demonstrated the best discriminatory performance among the models. Internal and external validation revealed that the AUCs of the osteoporosis model were 0.86 and 0.87, compared with the FRAX model (0.84 and 0.72)/OSTA model (0.77 and 0.66), respectively. The fracture model had high AUCs in the internal and external cohorts of 0.93 and 0.92, which were higher than those of the FRAX model (0.89 and 0.86). The survival model was also assessed and showed high reliability via internal and external validation (AUC of 0.96 and 0.95), which was better than that of the TNM model (AUCs of 0.87 and 0.87). Our models offer a solid approach to help improve decision making.

Keywords