A Generalizable Data Framework towards Precision Radiotherapy
Principal Investigator
Name
Jun Deng
Degrees
Ph.D.
Institution
Yale University
Position Title
Professor
Email
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCOI-626
Initial CDAS Request Approval
May 18, 2020
Title
A Generalizable Data Framework towards Precision Radiotherapy
Summary
In treating cancer patients with radiation therapy, different patients may have different responses to the same type of radiotherapy. Hence, it is critical to individualize the radiation treatment based on the patient's health data, clinical conditions, as well as response over time. The goal of the project is to develop a generalizable data framework that can support precision radiotherapy for individual cancer patients. Specifically, a deep reinforcement learning model will be built and validated with multimodal imaging data acquired during diagnosis, treatment and follow-up of individual patients. Harmonization of the medical imaging data with genetic and clinical data will create an invaluable repository of knowledge to draw from, while calling for new analytics. The developed data framework will provide critical clinical decision support for individualized radiotherapy.
By leveraging the wealth of data generated in the radiotherapy clinic, the project aims to develop a generalized deep reinforcement learning (DRL) tool for cancer risk stratification. Based on the DRL tool, an ensemble model will be built to analyze all the data types useful to patient outcome prediction. The model will be validated with independent datasets to ensure generalization. To account for information from multiple imaging modalities combined with treatment plans, a multimodal deep reinforcement learning (mDRL) model will be developed and trained with patient data stored in the electronic medical record system, as well as genomic information derived from blood and tissue specimens. The detection tool will be used in both lung cancer and colorectal cancer patients. Generalization to a variety of other cancers will be possible once the tools become available to the clinical research community. The ensemble model will allow integrated analysis of multiple data types recorded along the patient outcome trajectory, provide better discrimination between tumor phenotypes and superior predictive power. The framework will be designed to coordinate and synthesize various types of evidence and measurements into scores for the objective assessment and quantification of outcomes and endpoints. This strategy will ultimately provide novel patient re-stratification and support clinical decisions for highly individualized patient management.
By leveraging the wealth of data generated in the radiotherapy clinic, the project aims to develop a generalized deep reinforcement learning (DRL) tool for cancer risk stratification. Based on the DRL tool, an ensemble model will be built to analyze all the data types useful to patient outcome prediction. The model will be validated with independent datasets to ensure generalization. To account for information from multiple imaging modalities combined with treatment plans, a multimodal deep reinforcement learning (mDRL) model will be developed and trained with patient data stored in the electronic medical record system, as well as genomic information derived from blood and tissue specimens. The detection tool will be used in both lung cancer and colorectal cancer patients. Generalization to a variety of other cancers will be possible once the tools become available to the clinical research community. The ensemble model will allow integrated analysis of multiple data types recorded along the patient outcome trajectory, provide better discrimination between tumor phenotypes and superior predictive power. The framework will be designed to coordinate and synthesize various types of evidence and measurements into scores for the objective assessment and quantification of outcomes and endpoints. This strategy will ultimately provide novel patient re-stratification and support clinical decisions for highly individualized patient management.
Aims
1. Develop a deep learning model for cancer risk stratification
2. Build an ensemble modeling framework for patient outcome prediction
3. Validate the results with independent datasets for generalization
Collaborators
Enrico Capobianco, PhD, University of Miami
Harrison Zhou, PhD, Yale University