|
Abstract:
|
Treatment of head and neck cancers involves making multiple treatment decisions over time. Such treatment decisions must weigh the often competing goals of improving hard endpoints (e.g., 5-year survival) while minimizing toxicity. This interdisciplinary project describes results to design a novel approach for adaptive radiotherapy treatment of head and neck cancer which takes into account the patient's preference with regard to side effects. Specifically, we extend Q-learning, a type of reinforcement learning, to account for competing, multiple outcomes. We further extend the state of the art by taking into account both spatial data (such as medical images) and nonspatial data (such as demographics and toxicity) in the development of the treatment rules. We extract informative, relevant and non-redundant features from these high-dimensional patient data, while incorporating domain expert knowledge. We characterize patient similarity through data clustering that is robust to the effect of missing data and generate descriptors that further improve the reinforcement learning process. We illustrate our methodological development using a registry of patients with head and neck cancer.
|
