Explainable Machine Learning Explores Association Between Sarcopenia and Breast Cancer Distant Metastasis

Abstract

The impact of sarcopenia on the prognosis of breast cancer (BC) carries important clinical significance. However, there is no internationally standardized cut-off value for defining sarcopenia. This study proposed an explainable machine learning model for identifying risk factors of BC distant metastases and discussing the division of body composition cut-off values. Combining computed tomography (CT) image data of $11^{th}$ thoracic vertebrae (T11) and $4^{th}$ thoracic vertebrae (T4), a multi-objective optimized genetic algorithm was developed to select features and predict BC distant metastasis. Feature selection results were analyzed by Cox regression. The results showed that skeletal muscle index (SMI/T11) was a risk factor for predicting BC distant metastasis and an independent prognostic factor for distant metastasis-free survival (DMFS). A cut-off value of 21cm2/m2 for SMI/T11 was obtained by shapley additive explanations, in addition, The DMFS and overall survival (OS) of the low-risk group were significantly better than those of the high-risk group. The combination of multimodal data further confirms that sarcopenia is associated with poorer DMFS and OS in BC patients, and explores for the first time the issue of the cut-off values of sarcopenia and BC distant metastasis.

Keywords

• Breast cancer
• cut-off value
• distant metastasis
• multi-objective optimization
• multimodal
• sarcopenia