Development and validation of explainable machine learning models for the prediction of survival in patients with M1 breast cancer.

[BACKGROUND] The prognosis of patients with metastatic (M1) breast cancer is controversial, and the prognostic value of local therapy has not been well established. We aimed to develop and validate explainable machine learning (ML)-based survival models to predict overall survival (OS) in this population.

[METHODS] We retrospectively identified 10,214 female patients with histologically confirmed M1 breast cancer diagnosed between January 2013 and December 2018 from the Surveillance, Epidemiology, and End Result (SEER) database, each with a single malignant lesion. Patients with ambiguous or incomplete metastasis data were excluded. Candidate predictors included age; sex; laterality; American Joint Committee on Cancer (AJCC) Tumor, Node, and Metastasis stage; surgery of the primary site; breast subtype; estrogen receptor and progesterone receptor status; marital status; radiotherapy; chemotherapy; tumor grade; histology; and metastasis to the bone, brain, liver, and lung. Two time-to-OS prediction models-a neural network and a Cox proportional hazards model-were trained, internally validated, and externally tested in a cohort of 100 patients with M1 breast cancer from China. Model interpretability was assessed through global and individual feature importance analyses.

[RESULTS] In total, 10,314 patients were enrolled in the study. The median follow-up time was 42 months in the training dataset and 36 months in the test dataset. The deep learning network demonstrated greater stability and accuracy than did the Cox proportional hazards model in predicting patient survival, both on the internal test dataset (concordance index: 0.771 . 0.632) and in the external validation (concordance index: 0.782 and 0.650). Several important prognostic factors were identified by the deep learning model, including breast subtype, metastatic site, and surgery status. Surgery was associated with improved OS in patients with bone metastases selected after propensity score matching, with 5-year OS rates of 76.9% and 27.2% in the surgery and nonsurgery groups, respectively (P=0.001).

[CONCLUSIONS] We developed and externally validated ML models that accurately predict survival in patients with M1 breast cancer. Breast subtype, metastatic site, and surgery status were the most important factors for survival prediction in this population. Patients with non-triple-negative breast cancer and metastasis to the bone may benefit from surgery, while those with metastasis to the brain, lung, or liver may not.