Can we use AI to predict regional changes in vascular plasticity during cardiometabolic disease?

Code: BBSRC-DFA_2026_06

Primary Supervisor: Qianni Zhang
Email: qianni.zhang@qmul.ac.uk 
Institute: EECS

Secondary Supervisor: Neil Dufton
Email: n.dufton@qmul.ac.uk 
Institute: WHRI 

Abstract:

Tackling the growing global burden of Cardiometabolic disease (CMD) is a major clinical challenge of the 21st century which requires new approaches to improve risk assessment, diagnosis and therapeutics. The application of multi-omic technologies has produced large scale data in both pre-clinical and patient cohorts which have greatly enhanced our understanding of the underlying pathologies. One striking commonality between different etiologies of CMD is their impact on microvascular function during progressive disease. Our research has revealed that distinctive changes in the identity of endothelial cells (EC) occur early in pre-clinical models of CMD and are associated with detrimental outcomes such as fibrosis. We have termed the induction of these EC subpopulations fibrosis-associated EC (FAEC). However, the traditional diagnostic assessments of vascular dysfunction do not currently detect microvascular changes. We propose to develop a new AI pipeline to integrate longitudinal ultrasound echocardiology measurements with histological and single cell full spectrum flow cytometry analysis of FAEC. The AI pipeline will build on baseline deep learning networks to automate the extraction of key structural and functional parameters. By training this AI pipeline using extensive pre-clinical data from different animal models of CMD we aim to produce a new tool that can be applied to large-scale patient cohorts as a proof of concept. Taken together this studentship will amalgamate previously disparate imaging datasets to develop a comprehensive overview of spatial and functional changes associated with microvascular dysfunction in models of CMD.  

Lay Summary:

The dramatic global rise of metabolic disorders, particularly obesity and type 2 diabetes, has been central contributors to the marked increase in cardiovascular complications collectively termed cardiometabolic disease (CMD). Unfortunately, the patients who develop CMD are often not identified until they reach later stages of disease, such as the appearance of tissue scarring (termed fibrosis) which leads to irreversible tissue damage. One primary reason for missing CMD in patients is that the diagnostic analysis of clinical imaging, such as ultrasound echocardiology, is currently not sensitive to determine which patients are likely to progress towards cardiac fibrosis. Artificial Intelligence (AI) is already being successfully applied to echocardiography by offering operator assistance there by improving the capture, quality and ultimately interpretation of the data generated. However, we are still at the tip of the iceberg in incorporating its potential for predicting detrimental outcomes in patients.  

In this proposal we aim to use AI pipelines to bring together comprehensive profiling of a long-term high-fat diet model of mouse CMD. We will focus on developing AI analysis pipelines to integrate multiple echocardiology images and videos producing a comprehensive overview of the changes occurring in different regions of the heart over time. We will collate the AI analysis with multi-omic endpoint data including tissue pathology and single cell analysis (not routinely available in patients) to develop a predictive matrix of cardiac damage. The trained AI pipeline will then be assessed in open-access echocardiography databases to determine if we can enhance patient identification and help tackle the increasing burden of CMD. 

Aims and Objectives:

1. Segmentation of echocardiology videos and imaging data acquired longitudinally during mouse models of CMD to train an AI model based on video transformers or Mamba that can identify regional changes in cardiovascular function.  
2. Integration of multi-omic datasets, including immunofluorescence, single cell full spectrum flow cytometry, transcriptomics and metabolomics using a multimodal cross-attention fusion architecture to train an AI pipeline that creates a predictive matrix mapping to hepatic and cardiac outcomes. 
3. Assess conservation of microvascular dysfunction traits from the mouse models to both pig and human echocardiology datasets. Unsupervised domain adaptation and transfer learning will be applied to quantify cross-species transferability.