Approved Research

Objective assessment of depression from rsfMRI brain scans utilising white-box machine learning

Lay summary

Depression is the leading cause of illness, disability and death from suicide across the globe. There are no diagnostic tests or markers currently available. We cannot predict who will get depressed or when they will relapse. In patients that are unable to communicate this is especially difficult and in patients where there is diagnostic uncertainty.

Machine learning is where a computer learns how to predict an outcome of interest from data using algorithms. It has previously been shown to offer great potential in the diagnoses of many healthcare conditions. Using a proven machine learning approach developed by Smith, we plan to use resting brain scans (magnetic resonance imaging of brain activity whilst at rest) from a large existing dataset (the UK Biobank) to generate new algorithms that can identify participants with depression from healthy volunteers. One drawback of machine learning is that such algorithms are often uninterpretable to humans, meaning we do not know why the machine has made the decision it has. This can cause problems with both clinicians and patients accepting the decision. Crucially however, the machine learning approach we will use in this project is interpretable.

The project will run for 36 months. By developing algorithms that can accurately diagnose depression from brain scans, we will allow clinicians to offer treatments to patients more quickly and accurately. This will improve patient outcomes. If successful, the results obtained will be used to support an application for further, substantial funding to validate the technology, investigate other outcomes (such as response to treatments for depression) and see its translation into clinical and commercial use.