DOI
https://doi.org/10.25772/BEMC-J909
Author ORCID Identifier
0000-0003-3064-4088
Defense Date
2023
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Radiation Oncology
First Advisor
Jatinder R. Palta, PhD
Second Advisor
William C. Sleeman IV, PhD
Third Advisor
Elisabeth Weiss, MD
Fourth Advisor
Preetam Ghosh, PhD
Abstract
The proposed research aims to address the challenges faced by clinical data science researchers in radiation oncology accessing, integrating, and analyzing heterogeneous data from various sources. The research presents a scalable intelligent infrastructure, called the Health Information Gateway and Exchange (HINGE), which captures and structures data from multiple sources into a knowledge base with semantically interlinked entities. This infrastructure enables researchers to mine novel associations and gather relevant knowledge for personalized clinical outcomes.
The dissertation discusses the design framework and implementation of HINGE, which abstracts structured data from treatment planning systems, treatment management systems, and electronic health records. It utilizes disease-specific smart templates for capturing clinical information in a discrete manner. HINGE performs data extraction, aggregation, and quality and outcome assessment functions automatically, connecting seamlessly with local IT/medical infrastructure.
Furthermore, the research presents a knowledge graph-based approach to map radiotherapy data to an ontology-based data repository using FAIR (Findable, Accessible, Interoperable, Reusable) concepts. This approach ensures that the data is easily discoverable and accessible for clinical decision support systems. The dissertation explores the ETL (Extract, Transform, Load) process, data model frameworks, ontologies, and provides a real-world clinical use case for this data mapping.
To improve the efficiency of retrieving information from large clinical datasets, a search engine based on ontology-based keyword searching and synonym-based term matching tool was developed. The hierarchical nature of ontologies is leveraged to retrieve patient records based on parent and children classes. Additionally, patient similarity analysis is conducted using vector embedding models (Word2Vec, Doc2Vec, GloVe, and FastText) to identify similar patients based on text corpus creation methods. Results from the analysis using these models are presented.
The implementation of a learning health system for predicting radiation pneumonitis following stereotactic body radiotherapy is also discussed. 3D convolutional neural networks (CNNs) are utilized with radiographic and dosimetric datasets to predict the likelihood of radiation pneumonitis. DenseNet-121 and ResNet-50 models are employed for this study, along with integrated gradient techniques to identify salient regions within the input 3D image dataset. The predictive performance of the 3D CNN models is evaluated based on clinical outcomes.
Overall, the proposed Learning Health System provides a comprehensive solution for capturing, integrating, and analyzing heterogeneous data in a knowledge base. It offers researchers the ability to extract valuable insights and associations from diverse sources, ultimately leading to improved clinical outcomes. This work can serve as a model for implementing LHS in other medical specialties, advancing personalized and data-driven medicine.
Rights
© Rishabh Kapoor, Jatinder Palta
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
7-28-2023