Radiotherapy-induced cardiac and lung injury is a serious problem that is growing more critical with continuing efforts to escalate radiation dose to thoracic tumors. It is imperative that the mathematical tools be available to accurately predict such risk prior to treatment. Accurate prediction can help to mitigate the risk by appropriately altering the radiotherapy treatment plan. Inaccurate prediction, on the other hand, could have dangerous consequences by underestimating the probability of injury. Currently, popularly used parametric models can be inaccurate predictors due to certain assumptions. Parametric models assume that the mathematical behavior underlying radiation-response is dependent only on dose and is the same for all organs and injury endpoints. The hypothesis driving this project is that novel non-parametric models can more accurately predict radiotherapy-induced cardiac and lung injury than existing, popularly used, parametric alternatives. Injury data for this project will consist of dose data, SPECT-based functional maps, and relevant patient characteristics, from Duke University and the Netherlands Cancer Institute (NKI).
Specific aims supporting the hypothesis are: (1) Optimally fit commonly used parametric models to the input injury data. The model parameters will be fitted using maximum likelihood estimation, and parameter uncertainty will be estimated using Markov chain Monte Carlo simulation. (2) Develop novel non-parametric predictive models. These are powerful ensemble models that will combine the strengths of three component methods: linear discriminant, trees, and neural constructs. Construction of the component models will use novel techniques incorporating receiver operating characteristics (ROC) analysis to explicitly safeguard against overfitting to the input data. (3) Assess the robustness of the non-parametric models and compare their predictive accuracy to parametric models. A more robust model is less sensitive to the input dataset. Robustness assessment and model comparison will use novel techniques incorporating ROC analysis. ROC analysis will also be used to assess the predictive importance of using SPECT functional inputs. It is anticipated that the non-parametric models developed here will be significantly superior to existing parametric alternatives. They will be made publicly available via the Internet. The overall impact of this work will be to facilitate significant reduction in thoracic radiotherapy-induced injury via accurate prediction. ? ? ?
