This is a funded PhD position. The aim of this project is to provide explanations for 'black box' models of AI-based prediction of cardiovascular risk from retinal imaging. Reducing the cardiovascular disease burden is the core part of the NHS Long-term plan. Cardiovascular diseases cause a quarter of all deaths in the UK and are the most significant cause of premature mortality in deprived areas. It is the single most important area where the NHS can save lives. CVD is largely preventable through lifestyle changes and NHS action. The Plan highlights the need for early cardiovascular disease detection and that many people live with undetected, high-risk conditions (e.g., hypertension, raised cholesterol, diabetes), leading to a high overall cardiovascular disease risk. Only 45% of eligible people complete the NHS Health Check, and better ways to assess hard-to-reach populations are needed. Retinal photography that can be assessed using automated deep-learning algorithms is widely accessible (part of a routine eye check). The method has a good predictive value for cardiovascular disease risk but is limited by the deep-learning's ‘black box’ nature. The lack of interpretability in predictive models undermines trust in those models in health care. Black boxes also limit the clinical actionability of model predictions, which further undermines their usefulness to clinicians. The study aims to 'open' the 'black box'. This is an exciting opportunity to work across specialties linking data science and clinical practice.
OBJECTIVES:
1. To identify components of retinal photography used for cardiovascular prediction by deep-learning methods.
2. To identify the variables most driving the model predictions to provide explanations for black-box models at the global level (explaining the overall workings of a model) or the local level (explaining how the model reached a cardiovascular risk decision).
3. To apply methods of explainable machine learning to make the involved black-box models interpretable.
EXPERIMENTAL APPROACH:
UK Biobank (https://www.ukbiobank.ac.uk/enable-your-research/about-our-data) is a large UK dataset that holds both electronic health data and colour retinal photography which has already been successfully used for deep-learning-based cardiovascular risk stratification with our academic collaborators, Medi-Whale (Rim et al 2020 and 2021.). The proposed project will match items of retinal photography with clinical risk patterns (hypertension control, cholesterol levels, diabetes, etc.) and use them to apply methods of explainable machine learning. The student will apply Local Interpretable Model-Agnostic Explanations (LIME) explainability technique that provides easily understood explanations of which clinical factors contribute to each prediction. The methods benefit from ability to be linked into electronic health records. The student will also train a convolutional neural network on retinal fundus images to explain the ‘black box’ using Gradient-weighted Class Activation Mapping (Grad-CAM) at a global level. The successful PhD candidate will benefit from working with a multidisciplinary team with extensive experience in computer science, image processing, and medicine.
The candidate should have, or expect to have, an Honours Degree at 2.1 or above (or equivalent) in Mathematics, Statistics, Computer Science or related disciplines. It is essential to have background knowledge in machine learning, computer programming (e.g., Python), an interest in biomedical applications and a proactive approach to their work.
Please apply by sending your CV and covering letter to Dr Eduard Shantsila.
Continue with Facebook
