Predicting breast cancer among Hodgkin lymphoma survivors using radiotherapy dose distributions: Estimating risks of modern treatments from historic data

Cancer survival is increasing every year due to earlier diagnosis and advancements in cancer treatment, implying that an increasing number of people with a history of cancer are alive each year, who are at risk of developing second primary cancers. The risk of second cancers is associated with previous cancer treatment, e.g., breast cancer (BC) is an important late complication for survivors of Hodgkin lymphoma (HL) who were treated with radiotherapy to the chest. Predicting second cancer risk is important, but a difficulty faced when developing a prediction model is the long induction period between treatment for the first cancer and diagnosis of the second malignancy. Therefore, prediction models are based on patients treated decades in the past. Because treatment strategies change over the years, the resulting models may not work well for more recently treated patients. Published studies examining the association between radiotherapy and BC risk used relatively crude metrics of radiation exposure of the breast. The aim of this thesis was to improve on this by using more detailed information on radiation exposure, in particular the radiation dose distribution in the organ at risk. In Chapter 2, we examined the statistical properties of multiple methods of using dose distribution data to estimate dose-related risks from matched case-control studies. Simulation studies showed an overestimation of more than 30% of the excess relative risk per Gy (ERR/Gy) with realistic sample sizes of 75-150 cases. Firth's bias correction reduced but did not entirely eliminate bias. In Chapter 3, we summarized the spatial dose distribution of participants of a nested case-control study of BC among 5-year HL survivors using dose-volume histograms, and used this to study the joint effects of radiation dose and irradiated volume on BC risk among HL survivors. The data showed a linear relationship between mean breast dose and BC risk (excess odds ratio/Gy 0.19) after adjusting for duration of post-radiotherapy intact ovarian function, without a significant curvature. In Chapter 4, we described the development of our absolute BC risk prediction models. For this, we used the above-mentioned case-control study nested in a cohort of Dutch HL survivors. We developed models using overall mean breast dose and dose to ten breast segments in addition to several other factors. Based on internal validation, all models were well calibrated with observed/expected ratios between 0.85 and 1.01, and all yielded modest cross-validated AUCs around 0.57. In Chapter 5, we outlined the clinical implications of the BC risk prediction models and externally validated them in childhood cancer survivors from the US Childhood Cancer Survivor Study cohort. The ERR/Gy was 0.26 for the mean breast dose model and 0.14 for the breast segment-specific model, and BC risk was 3-fold higher in the upper outer quadrant compared to the upper inner quadrant. Models had a good overall calibration (observed/expected ratios of 1.22 and 1.12) and moderate discriminatory performance (AUC 0.68), with no difference between the breast segment-specific model and the simpler mean breast dose model. In Chapter 6, we compared breast radiation dose distributions for 113 patients treated with chest radiotherapy since 2006 to historic treatments given in 1965-2000. Radiation doses and volumes decreased substantially between the time periods, and most dose-volume metrics continued to decrease over time since 2006. Based on segment-specific dose, median absolute BC risk 25 years after treatment for a high risk patient decreased from 21.4% to 7.1% for historic vs. recent treatments, respectively, and from 3.2% to 1.0% for a low risk patient. Across recent treatments, the breast segment-specific model predicted 14.9% and 18.0% higher risk than the mean breast dose model for a low and high risk patient, respectively.