Learning to Trust AI Models in Cardiology

Lead Supervisor
Dr Andrew King
Reader in Medical Image Analysis
Biomedical Engineering, Biomedical Engineering and Imaging Sciences, FLSM, King’s College London
andrew.king@kcl.ac.uk

Co-supervisor
Professor Reza Razavi
King’s College London

Project Details

Heart disease is the number one cause of death worldwide and is the most common cause of heart failure. Many new technologies have been introduced in recent years, resulting in a huge amount of data upon which clinical decisions can be based. However, despite this, currently decisions about the treatment and management of heart disease are typically based upon a tiny proportion of these data.

This project aims to develop an artificial intelligence (AI) decision-support tool that can use big data to assist cardiologists in making better decisions about treating heart failure patients. Heart failure can have a number of causes, including coronary artery disease, cardiomyopathy and congenital heart disease. Treatment or management options can include drug treatment, implantation of a device such as a pacemaker, simple lifestyle changes or surgery. It is essential that the right diagnosis is made to avoid incorrect treatment decisions, and this will be the primary purpose of the AI decision-support tool.

The project will contain three key novelties aimed at promoting the clinical translation of the tool:

  1. Big data
    The decision-support tool will be able to exploit the vast amount of information available in different imaging modalities as well as non-imaging data such as electronic health records (EHRs). We will develop AI techniques based upon the latest deep learning models to learn useful knowledge representations from MRI data (cine, late gadolinium enhancement, T1 shMOLLI) and ultrasound data. We will also make use of Natural Language Processing tools such as CogStack to extract meaningful and robust representations from the free text EHRs. Finally, we will develop deep learning architectures that can make decisions using these multiple data sources, and make them robust to noisy, corrupted or even missing data.
  2. Explanations
    We will develop techniques based upon the latest explainable AI research to enable the decision-support tool to explain its decisions using concepts that are well understood by cardiologists. This will involve using deep learning architectures such as variational autoencoders to make links between high-level concepts provided as prior knowledge by cardiologists, and lower-level (temporal) image features.
  3. Trust
    The explanations, as well as the primary output of the tool, will be associated with confidence measures, enabling cardiologists to make informed decisions about whether to accept the AI tool’s recommendation or not. To this end, we will investigate trust measures based upon information-theory metrics computed from the knowledge representation space to quantify the reliability of explanations and decisions.

This is a strongly translational project and the two main supervisors are a computer scientist who specialises in AI and a cardiologist. The student would have the opportunity to work alongside AI experts to develop the appropriate technical skills and expertise. They would also work closely with NHS cardiologists to learn about how they make clinical decisions and what language they use when discussing cases. This engagement will be essential to achieve the aim of developing a decision-support tool that can explain its decisions using cardiologist-interpretable concepts and back this up with evidence from the data.

Datasets

We will make use of two large datasets in this research.

  1. UK Biobank: This dataset currently contains approximately 50,000 cases (this will eventually rise to 100,000) with cardiac MRI data together with a wide range of non-imaging data fields. The subjects are a cross-section of the UK population, so most are healthy but there are cases diagnosed with heart disease and heart failure and for these subjects diagnostic information is available. The primary supervisor has an existing approved UK Biobank project (number 17806) which the proposed work falls within the scope of. All UK Biobank volunteers have already consented to use of their data for approved research projects. Data are held on a secure file server at KCL with limited user access.
  2. KCL/GSTT Bioresource: We also have access to over 25,000 retrospective cardiac MRI scans from patients scanned at Guy’s and St. Thomas NHS Foundation Trust (GSTT) with heart disease (of which approximately 5,000 are heart failure patients). This number will increase by approximately 5,000 per year (1,000 with heart failure). Opt-out ethical approval is in place for research use of all retrospective and future data within the bioresource. Through the London Medical Imaging & Artificial Intelligence Centre for Value-Based Healthcare technological infrastructure is being developed to enable easy access to deidentified imaging data from this bioresource. Linked patient EHR data will be held on a secure GSTT server with the capability to upload and train/evaluate AI models.

Keywords

AI, cardiac, explanations, trust, EHR