Prediction of Side Effects from Breast Radiation Therapy - Integration of Clinical and Genomic Data.

OBJECTIVES: To compare the ability of different machine learning models to predict the risk of side effects in patients with breast cancer undergoing radiation therapy. METHODS: Data from the multicentre REQUITE cohort (n=2067) was analysed retrospectively. Side effects (8 endpoints) were assessed 24 months post radiation therapy. Clinical, treatment and genetic data were available. Predictive performance of 12 machine learning models was assessed using Area Under the Receiver Operating Characteristic Curve (AUC-ROC) and Area Under the Precision-Recall Curve (AUC-PR), with n-repeated k-fold Cross Validation (CV). RESULTS: Random Forest achieved the best performance using clinical data alone (mean AUC-ROC=0.84, 5 × repeated 10-fold CV). Combining clinical and Single Nucleotide Polymorphism (SNP) data yielded the highest overall performance with Logistic Regression (mean AUC-ROC = 0.92, 5 × repeated 10-fold CV), as the highest followed by ensemble tree classifiers (mean AUC-ROC = 0.91, 5 × repeated 10-fold CV). The best prediction was obtained for arm lymphedema, followed by breast oedema and nipple retraction. CONCLUSION: In this large multicentre cohort, radiation therapy side effect prediction was highly endpoint-dependent, with lymphedema emerging as a particularly robust outcome. When combining clinical and SNP data, Logistic Regression performed comparably to more complex approaches, while offering greater transparency. External validation and cost-benefit evaluation are required, but our findings provide a comparative benchmark and support the feasibility of targeted prediction of breast radiation therapy side effects.