Approved Research

Multi-view unsupervised generative models for studying heterogeneous diseases.

Lay summary

The overarching aim of this research is to measure and quantify how much an individual differs from healthy individuals across different types of medical data. This is motivated by several factors. Firstly, there is often lots of variability, for example in disease severity, within a disease cohort which is not captured using existing methods. Secondly, different types of medical data capture different information about a patient. As such it is useful to explore methods of combining different types of data. Thirdly, there is a general shift, as we learn more about complex diseases and develop a wider portfolio of treatments, towards personalised medicine. This research will provide a measure of deviation from the healthy pattern at the individual level which could, for instance, be used in treatment planning.  

This work combines the topics of normative modelling, generative modelling and multi-view methods. Normative modelling is a statistical modelling approach used to measure how much diseased individuals deviate from a healthy population. Multi-view methods are machine learning methods which combine multiple different types of data into one framework with the aim of incorporating information from different data sources or finding relationships in the data. Both normative modelling and multi-view methods have been applied to the study of neurodegenerative diseases. However, there has been limited research combining the two research areas. Generative models can be used to create new data. This is useful when some data types are missing, a common problem in medical datasets.

This project covers both new methodological development and application to the study of a number of neurodegenerative and cardiovascular diseases. We expect the project to last up to 3 years, including the time required to write up and publish the research findings.

Once trained, normative models require little compute time to provide a measurement for a patient. They could easily be implemented in a clinical setting and provide medical practitioners with a measure of abnormality without having to rely on a previous disease diagnosis. This could provide insight into a patient's condition without bias towards a particular disease and help derive patient specific treatment or flag patients which require closer monitoring.