A new machine learning model has been developed to predict the risk of secondary cancers in patients who have undergone radiation therapy. These secondary cancers can appear years later due to the effects of ionizing radiation. The model, a Random Forest Regressor, uses data from clinical, pathology, and genomic sources—such as cancer type, age, radiation dose, tumor grade, and key mutations like TP53 and BRCA1/2.
Using cancer registry data, the model achieved strong performance with an MSE of 0.002 and R2 of 0.98, outperforming traditional prediction methods. The most important predictors were radiation dose and age at exposure, followed by TP53 mutations. The model also identified high-risk groups, such as Hodgkin lymphoma survivors (with a 25.4 per 10,000 risk of secondary breast cancer) and breast cancer survivors (with a 15.2 per 10,000 risk of secondary lung cancer). A non-linear dose-response pattern was observed, showing a 3.5-fold increased breast cancer risk at doses between 30–40 Gy for Hodgkin survivors.
This machine learning framework could support personalized follow-up and surveillance in cancer survivors. However, the study notes that larger, multi-center datasets and imaging data integration are needed to further improve accuracy.