ESTRO 2025 - Abstract Book

S3915

Radiobiology - Normal tissue radiobiology

ESTRO 2025

1418

Digital Poster A novel method for defining subgroups of breast cancer patients with different risk of fibrosis after radiotherapy Ahmad Sami 1,2 , Marlon R. Veldwijk 1,2 , Frank A. Giordano 1,2 , Jenny Chang-Claude 3 , Petra Seibold 3 , Carsten Herskind 1,2 1 Department of Radiation Oncology, Universitätsmedizin Mannheim, Medical Faculty Mannheim, Heidelberg University, Mannheim, Germany. 2 *, DKFZ-Hector Cancer Institute at the University Medicine Mannheim, Mannheim, Germany. 3 Division of Cancer Epidemiology, German Cancecr Research Center (DKFZ), Heidelberg, Germany Purpose/Objective: Multiple risk factors with low to modest effect sizes are considered to contribute to patients’ individual risk of late toxicity after radiotherapy. The purpose of the present study was to test the alternative hypothesis that some factors are important only within certain subgroups characterized by different functional pathways (1) and to assess its potential for predictive modeling. Material/Methods: DNA, CD4 + radiation-induced lymphocyte apoptosis (RILA) data, and long-term (10.3-12.8 years) follow-up data, were available from 238 patient in the German ISE breast cancer cohort (2). 71 patients (29.8%) developed moderate-severe fibrosis. Three candidate SNPs selected from a small mass array pilot study were genotyped by PCR in these 238 patients. Predictive modelling was performed using partition analysis (PA) and ensemble Machine Learning (ML) modeling. The identity of the SNPs and details of the ML models will be revealed at the ESTRO meeting. associations with breast fibrosis. Using these features together with RILA, PA identified six subgroups. BMI and the two SNPs were significantly associated with fibrosis (one of the SNPs was protective) in the RILA-low subgroup (which was associated with increased risk) but not the RILA-high subgroup (associated with reduced risk) whereas the opposite was true for HTN. None of the features were significant in the complementary RILA subgroups, strongly supporting the subgroup hypothesis. The six subgroups could be assigned to three risk groups (low, intermediate, high) spanning a five-fold difference in risk (p<0.0001). Two types of ML modeling including cross validation were performed. Combining the two ML models yielded three risk groups which were highly correlated (p<0.0001) with those from PA. Based on the five-feature predictive ML model, 80% of the patients were assigned to high (13.7% with 83.3% risk) or low-risk groups (66.7% with 17.1% risk), yielding AUC=0.735 in ROC analysis. Conclusion: The results strongly support the subgroup hypothesis and the robustness of the combined ML model. However, at present no other cohort with long-term follow-up and CD4 + RILA data exists worldwide for external validation. Based on just five features, the combined ML model might potentially spare more than a third of the patients from developing fibrosis by offering an alternative treatment (e.g. partial-breast radiotherapy) to high-risk patients. The correlation with PA allows mechanistic interpretation and provides scope for further improvement. Results: Two SNPs (SNP A and SNP B ) and two clinical parameters (BMI and hypertension, HTN) showed significant

Keywords: Fibrosis, prediction, machine learning

References: 1. Herskind et al., Cancer lett. 382:95-109, 2016. 2. Veldwijk et al. Clin. Cancer Res. 25:565-72, 2019.