Predicting early respiratory-related mortality after lung cancer radiotherapy

Abstract

In this thesis, we aimed to address the time-consuming manual preparation of radiotherapy (RT) data, which poses a barrier to large-scale dose-response studies, by developing automated methods for this task. Using data from multiple hospitals prepared with these methods, we also aimed to develop risk models for early respiratory-related mortality following RT for non-small-cell lung cancer. The thesis encompasses five papers. In Paper I, we investigated the feasibility of automatically collecting available RT data from four hospitals. Patient age and absorbed dose to the lungs were combined with manually collected outcome data to estimate a logistic risk model for early respiratory-related mortality. In Paper III, we developed a semi-automated workflow for efficient preparation of additional RT data, including data cleanup, quality controls, and automatic segmentation of the organs at risk (OARs) heart, proximal bronchial tree, and esophagus with an in-house deep-learning segmentation model. In Paper IV, we performed quality assurance of this segmentation model, trained with cases from a single hospital, to assess its suitability for generating reliable dose data across all four hospitals. Quality controls focusing on the autosegmentations were also developed. In Paper II, we simulated the impact of delineation errors on the estimated OAR doses as well as the impact of OAR dose errors on estimated dose-response relationships, informing data preparation choices for Paper V. Finally, in Paper V, we evaluated candidate predictors of early respiratory-related mortality among patient age and absorbed doses to the lungs, heart, proximal bronchial tree, and esophagus. The key predictors were used to estimate a risk model by combining cause-specific Cox regression and the Aalen-Johansen estimator. In conclusion, our automated methods enabled multi-hospital preparation of RT data for dose-response modeling. Using these data, we identified patient age and absorbed dose to the lungs as key predictors of early respiratory-related mortality and developed a risk model based on these predictors.