In a large-scale retrospective analysis, the MACHINE learning model accurately predicted survival outcomes in patients with breast cancer bone metastases while identifying individuals at high risk of early death.
Researchers analyzed data from 23,723 patients newly diagnosed with stage IV breast cancer and bone metastases between 2010 and 2020 and compared multiple machine learning approaches for predicting survival.
The model was then evaluated in an independent multicenter cohort of 230 patients.
LightGBM emerges as a leading prognostic model
Among all algorithms evaluated, the Light Gradient Boosting Machine (LightGBM) model showed the most consistent prognostic performance, with decision curves and calibration analysis supporting the model’s potential clinical utility.
The study also focused on early mortality, defined as death within three months of diagnosis.
Approximately 16% of patients in the population-based cohort experienced early death within this period.
Although the stacking model was best at distinguishing between patients who experienced early death and those who did not, LightGBM had the most balanced overall performance across measures of accuracy, precision, specificity, calibration, and clinical utility.
Sensitivity analyzes excluding surgery, radiotherapy, and chemotherapy variables also showed that LightGBM yielded acceptable predictive performance.
According to the researchers, this suggests that the model may use information available at the time of diagnosis to support risk assessment.
Surgery is associated with improved survival
The researchers also investigated the association between primary tumor surgery and survival outcomes.
Before matching, surgery was associated with improved survival compared with no surgery, but the survival benefit remained significant after propensity score matching.
Similar findings were observed in the external cohort.
However, the researchers stressed that treatment response, performance status, frailty, and metastatic burden were not available, meaning that residual confounding and selection bias could not be ruled out.
Therefore, observed associations should be interpreted with caution and not as evidence of causation.
Future prospective studies incorporating broader clinical data will be needed to validate these models and clarify how they can support decision-making in routine oncology care.
reference
Fang J et al. Machine learning-based prognosis and early death prediction in patients with de novo stage IV breast cancer with bone metastases: SEER database and multicenter retrospective study. Sci Rep. 2026;DOI:10.1038/s41598-026-62142-w
Featured image: romaset by Adobe Stock
