Project 1: Learning to Stratify Psychotic Disorders using Neuroinflammatory Biomarkers and Brain Imaging
Psychotic disorders are costly conditions, with billions spent on healthcare and over £18 billion annual economic impact in the UK. Early diagnosis improves recovery and reduces burden, but psychosis patients present diverse symptoms, characteristics, and treatment responses. Classifying patients into subgroups enables identifying tailored treatment options to improve outcomes.This project aims to predict which high-risk individuals will develop psychosis by analysing blood biomarkers combined with brain imaging data. We also seek to understand factors contributing to variability among psychosis patients to facilitate earlier intervention and personalised treatments.
Our analysis integrates inflammatory biomarker measurements and brain MRI scans. Advanced statistical techniques will study connections between biomarkers and brain functioning. AI and computer science methods will create predictive models classifying patients based on biomarker levels and brain imaging features. This classification can guide understanding of which treatments are most effective for different subgroups. This novel approach offers a personalised, proactive strategy for early psychosis intervention by accounting for disorder heterogeneity. Integrating blood biomarkers with brain imaging has the potential to revolutionise treatment outcomes through tailored care from the earliest stages.
Project 2: Developing mathematical models for seizure forecasting in epilepsy
Epilepsy is a condition that affects the brain, causing repeated seizures. Epilepsy affects around 630,000 people in the UK. Epilepsy can have a big impact on people's lives. It puts people at greater risk of injury and premature death, and it can affect people's social life, work life, and mental health. People living with epilepsy say that the unpredictability of seizures is one of the biggest problems. We believe that technology can predict when an epileptic seizure is likely to occur. You could think of it as being similar to a weather forecast. Seizure forecasts could provide people with epilepsy with a forecast about their risk of having a seizure in the near future. This forecasting would rely on us collecting a range of information about people living with epilepsy. Information such as sleep quality, stress levels and what medication they take. These are required as they are well known as ‘triggers’ which increase the likelihood of a seizure happening. All your data would be used as input to our system. Our system would then crunch the numbers, using advanced mathematics, and output an estimation of how likely you are to have a seizure. Over time, the estimations are likely to get more accurate. As the system gets to know the person living with epilepsy.
Project 3: A pseudospectral method for predicting patient response to depression medication from fMRI data
Depression is a major mental health disorder which affects about one in twenty people. It can be treated, often through a combination of psychotherapy and medication, but not everyone responds well to every drug, so there is a great need for methods which would help us to guide treatment. We have data from a drug trial which compared the effects of a standard antidepressant with a natural hallucinogenic compound, including brain scans using functional magnetic resonance imaging (fMRI). In previous research we have transformed these images into networks representing the flow of information in the brain, and shown that the effect of each drug can be seen in changes to network structure. We now wish to build on this work by applying mathematical techniques (such as computing the pseudospectra) which can tell us how a system will respond to perturbations - for example, to a medicine. Our aim is to develop a way of helping clinicians and patients to decide which treatment would be best in each case. We also hope this work will improve our understanding of how different drugs affect our mental states, and open up new avenues of research for brain imaging to inform psychiatry.
Project 4: Directed forgetting in human brains and artificial neural networks
Working memory describes the ability to remember and manipulate information that is no longer available in the environment. It provides a flexible mental workspace that scaffolds most higher mental functions. A key feature of human working memory is its limited capacity: we can only hold on to a small amount of information at any given moment. To make the most of this precious resource, a mechanism is needed that rapidly removes memories when they are no longer needed. Remarkably, however, humans often struggle with this. Difficulties in goal-directed forgetting are particularly have been linked with the risk of developing some forms of mental illness and may also contribute to age-related memory deficits. However, the mechanisms that underpin directed forgetting remain unknown. To fill this gap, we plan to study how artificial neural networks from machine learning form and remove memories in the same tasks that are used to study humans. These networks provide a unique window to reveal information processing mechanisms, as the activity of every single neuron is known and can be manipulated. We will leverage this virtue to reveal how neural circuits implement directed forgetting and develop testable new predictions for aberrant memory removal in mental illness.
Project 5: Developing machine learning probabilistic models towards improving the diagnosis of acute compartment syndrome
Acute compartment syndrome (ACS) is an emergency medical condition in which a high-impact trauma such as a fracture or crush injury causes pressure build-up within the limb which obstructs blood flow to the muscles of the limb. As the limb is dying slowly due to lack of oxygen and nutrients, the surgeon is attempting to diagnose the condition accurately. Due to lack of objective methods, clinicians depend upon subjective methods of diagnosis such as pain experienced by the patient upon stretching the affected limb. Such subjective assessments can be difficult even for experienced surgeons and can have serious consequences to patients such as having them to undergo a highly invasive surgery as a result of misdiagnosis or overdiagnosis. The current project aims to develop ML-driven ACS diagnostic tools that will help clinicians make informed decisions. We aim to develop ML models based on the ACS patient data that we have obtained in order to decipher the most sensitive/specific parameters that will aid in accurate diagnosis. We then aim to test the accuracy of our ML models against the current diagnostic methods. Finally, we aim to develop the ML models into a web application for wider dissemination among surgeons.