Novel method for risk stratification of radiation-induced breast fibrosis: subgroup hypothesis verified by machine learning.

Breast fibrosis (BF) after radiotherapy remains one of the most dreaded late toxicities in breast cancer care, yet multiple additive predictors struggle to capture its underlying biological complexity. Radiation-induced lymphocyte apoptosis (RILA) has recently been associated with the risk of fibrosis more than 10 years post-RT. Here, we show that a combination of five independent factors, RILA, two SNPs in the CTGF and NBS1 genes, and two clinical variables (body-mass index and hypertension) exhibits several important interactions. Partition analysis identified six partly nested subgroups, which could be consolidated into three clinically meaningful risk groups. Machine-learning modelling verified and refined these groups, demonstrating a five-fold variation (17-83%) in BF risk with an AUC = 0.735 in ROC analysis using only these five features. Our study provides proof-of-concept that a biologically realistic subgroup-based approach sharpens predictive performance and may enable clinical identification of a subgroup of breast cancer patients highly susceptible to BF.