A significant challenge in lung cancer radiation therapy (RT) is respiration-induced tumor motion, which hinders sufficient delivery of curative doses to target volumes. Although modern tumor motion management strategies for positron emission tomography/computed tomography (PET/CT)-guided RT are becoming more available, those techniques have yet to be fully incorporated into clinical practice. This is mainly because not every patient will benefit from a costly and lengthy motion-managed PET/CT scan due to high intra-patient and inter-patient variability of respiratory patterns. The objective of this project is to bridge the knowledge gap of which motion management method would best benefit an individual patient. This project will develop a new decision-making paradigm, in which machine learning techniques will be developed to characterize respiratory motion patterns and combine them with other diagnostic factors to predict the benefits from motion management methods for each individual patient. A decision-analytic cohort model will be developed to compare and evaluate the cost-effectiveness of the new decision paradigm and the traditional population-based radiation oncology practice of motion management based on our existing database of respiratory traces from more than 3,000 patients. While specifically applied to decisions surrounding respiratory motion management, the developed decision paradigm can be generalized and applied to other real life decision analysis problems.
This award supports fundamental research in data mining/machine learning and decision analysis, which will provide needed knowledge for the development of tools for effective management of patient-specific tumor motion. The modeling effort in this project will 1) establish a new mathematical foundation for supervised multivariate sparse variable selection and prediction to discover complicated multivariate relationships among high-dimensional variables; 2) construct a general integrated validation framework to rigorously test the cost-effectiveness of patient-specific health interventions. The new multivariate sparse variable selection and prediction approach can be used to build an interpretable prediction model, handle high-dimensional data with a low sample size, avoid under-shrinkage effect, and incorporate structured group selection. The cost-effectiveness analysis framework integrates the outcome of prediction model, the treatment effect and survival outcome model. This modeling aims to quantitatively estimate long-term cancer survival outcomes from improvement in patient-specific planning of radiation dosing by selective motion control.
