Skip to navigation Skip to main content Skip to footer

Approved Research

Optimal patient-level eye and systemic disease risk prediction from heterogeneous, multi-modal data sources using machine learning

Principal Investigator: Mr Justin Engelmann
Approved Research ID: 91570
Approval date: December 22nd 2022

Lay summary

The last decade saw rapid advancements in machine learning, also known as artificial intelligence (AI). Such AI models can learn from vast quantities of data and make predictions that often come close to or match the performance of human experts. In some cases, AI models can even do things that human experts cannot. For example, AI models can tell the sex of a person from a picture of the eye (specifically of the retina, a layer of tissue at the back of our eyes that allows us to see). We did not know that this was possible, and humans still cannot do it nearly as well as AI models.

Using AI models to support doctors could allow them to make more accurate diagnoses, and to identify patients at risk which they can then examine more often. This is very important as for many diseases, early detection is key to better outcomes. However, there are many open questions and challenges that limit the usefulness of AI models in healthcare. Our research aims to develop better and more practical AI models by combining expertise from AI, biomedicine, and clinical practice.

Despite the name, AI models are not intelligent in the way that humans are: For example, if a model has been trained to predict someone's risk of a heart attack from their age, sex, blood pressure and BMI, then this model cannot be used if we do not know their blood pressure. The same model also cannot consider information that it was not trained on, for example whether the patient smokes or not.

We aim to overcome this challenge by developing a method to train AI models for different contexts, so the doctor always has an AI model that uses all available information. This will allow us to investigate what information is most important for making accurate predictions. Our results can then inform what tests or scans a doctor should do and which ones are less useful. Furthermore, we want to explore what diseases can be detected and predicted from the images of the eye mentioned above and how we can use them most effectively.

If successful, our research would provide better AI models for healthcare and in turn improve our health by allowing doctors to make more accurate and earlier diagnoses. This is initially a 3-year PhD project, but we hope that it will form the foundation of a scientific career.