Development of machine learning models integrating multi-modal medical data for identification of disease therapeutic targets from the UK Biobank
Approved Research ID: 80521
Approval date: April 20th 2022
The wealth of clinical imaging and non-imaging clinical measurements and risk factors in the UK Biobank offers a unique opportunity to investigate the health status of organs both separately and simultaneously within the same individuals, however the methods for such analysis are yet not well developed.
Recent developments in machine learning, especially deep learning algorithms, to learn relevant features from raw imaging data, constitute a paradigm shift opening exciting opportunities to investigate medical imaging at scale. Deep learning approaches have been shown to yield results of comparable accuracy to human experts in various clinical applications.
While those ground-breaking solutions offer automation in various areas of healthcare, there has been little work exploring deep learning methods to predict disease progression at a patient level using multi-modal i.e. imaging and non-imaging studies in large data sets (e.g. UK Biobank).
We will develop machine learning methods to extract useful features primarily from the medical imaging available in the UK Biobank to model health outcomes. The new methods will include imaging, non-imaging data, environmental and genomic data to build personalised prediction models, and the UK Biobank data will also be used for evaluation and testing the new methods.
We anticipate that the project is mostly devoted to a new methodology development, which will be validated and tested to extract new multi-modal features from the resources available in the UK Biobank.
While the proposed project is motivated by methodology development, the wealth of multi-modal data in the UK Biobank, the number of cardiometabolic diseases cases and advances in machine learning methodologies, together opens an opportunity to create and test potential tools to model e.g. cardiometabolic disease progression. Therefore, as an example, we will be developing a prediction model using both imaging and clinical measurements related to microvascular complications.