Machine learning-based prognostic model for metastatic breast cancer and its interpretability: a multicenter...

Area of Science:

• Oncology
• Biostatistics
• Machine Learning in Healthcare

Background:

• Metastatic breast cancer (MBC) prognosis faces challenges with current scoring systems' accuracy and machine learning (ML) model interpretability.
• Accurate prognostic models are crucial for improving survival prediction and guiding individualized treatment strategies in MBC.

Purpose of the Study:

• To develop and validate a precise prognostic model for predicting overall survival (OS) in MBC patients.
• To leverage ML techniques on the Surveillance, Epidemiology, and End Results (SEER) database for enhanced prognostic evaluation.

Main Methods:

• Utilized data from 1,385 MBC patients from the SEER database, split into training (1,035) and internal validation (350) cohorts.
• An external validation cohort of 73 patients from Jiaxing Women and Children's Hospital was established.
• Key OS predictors were identified via multivariate Cox regression; four ML algorithms, including random survival forest (RSF), were employed to construct models.

Main Results:

• The RSF model demonstrated superior performance with a C-index of 0.723 in training and 0.727 in internal validation cohorts.
• RSF outperformed other models in AUC and Brier scores, proving effective for 1-, 3-, and 5-year OS prediction.
• External validation showed a stable C-index of 0.685; SHapley Additive exPlanations (SHAP) identified triple-negative breast cancer (TNBC) and brain metastasis as significant mortality risk factors.

Conclusions:

• The developed RSF prognostic model offers excellent predictive accuracy and interpretability for MBC survival.
• The model's ability to identify high-risk factors facilitates personalized treatment planning for MBC patients.
• This ML-driven approach enhances prognostic assessment, potentially improving clinical decision-making for metastatic breast cancer.