Rutgers Cancer Institute of New Jersey New Brunswick, NJ
S. Wei1, Y. Zhang2, S. Sowmiyanarayanan3, Z. Abou Yehia1,4, I. Jan4, N. J. Yue1, B. G. Haffty4, and K. Nie2; 1Rutgers Cancer Institute of New Jersey, New Brunswick, NJ, 2Department of Radiation Oncology, Rutgers Cancer Institute of New Jersey, New Brunswick, NJ, 3Emory University, Atlanta, GA, 4Rutgers Cancer Institute of New Jersey, Department of Radiation Oncology, New Brunswick, NJ
Purpose/Objective(s): Acute radiation dermatitis (RD) is a notable side effect in breast cancer patients undergoing radiation therapy. This study aims to predict such adverse effect through an integrated machine learning model that synthesizes clinical data, mammographic radiomic features, and genetic mutations identified by single nucleotide polymorphism (SNP) testing. Materials/
Methods: In this retrospective study, data from 174 breast cancer patients treated with radiation following breast-conserving surgery or mastectomy between 2004 and 2016 were evaluated. Genetic profiling included 30 genes, sequenced from peripheral blood samples. Extracted genomic DNA was annotated using Illumina Variant Studio Software and verified with ClinVar database. Patients were stratified based on germline genetic status into four groups: non-carriers, carriers of benign or likely benign variants, variants of uncertain significance (VUS), and pathogenetic variants. Pre-surgery contralateral mammograms were processed to extract 94 radiomics texture features using open source software. Clinical parameters clinical parameters such as race, age, BMI, TNM staging, total radiation dose, and hypofractionation were also recorded. RD of grade 2 or above were identified as positive cases in this work. A linear support vector machine (LSVM) model was trained with 5-fold cross-validation, to identify significant predictive features from the compiled clinical, radiomics and genomics data. Results: The LSVM model with genetic data integration pinpointed 3 genetic markers (CDH1, PTEN, TGFBRAP1), 4 radiomic features, and 2 clinical features (hypofractionation and race) as significant predictors of radiation side effects, achieving an area under the ROC curve (AUC) of 0.75 (± 0.04). The inclusion of mammographic patterns offers additional insight into correlation between germline genetic status and the associated toxicities. Conclusion: This study demonstrates the potential of a radiogenomic approach to predict acute radiation therapy side effects in breast cancer patients. The preliminary findings underscore the need for further investigation into the underlying mechanisms linking mammographic appearance and genomic features to side effects, facilitating personalized treatment strategies to mitigate these adverse outcomes.
