Automated Personalised 4D Heart Modelling for Disease Prediction

Generative statistical models of cardiac anatomy and function have a wide range of applications such as disease diagnosis, personalised medicine, generation of population and sub-population cohorts for in silico trials, etc. Geometric deep learning methods for cardiac anatomy have shown promising results for reconstructing the 3D biventricular cardiac anatomy shapes conditioned on population characteristics and predicting the 3D shape deformations between heart contraction and relaxation. Until now, research has focused on specific states within the cardiac cycle; however, the whole cardiac motion pattern plays a crucial role in determining the underlying pathologies.

To address this significant gap in the current research, our aim is to develop a novel multi-modal model for the complete cardiac motion over the 3D cardiac anatomy from the standard clinical cardiac magnetic resonance imaging (MRI), using deep learning-based approaches. In addition to the information captured by cardiac MRI, the proposed model will also incorporate multi-modal information including individual's demography such as age, sex, ethnicity, body mass, etc. and electrophysiology data to accurately model and quantify their relationships. We intend to develop the proposed approach on the large and diverse UK Biobank population, in order to investigate the differences in the different data modalities. The envisaged final 4D model of cardiac anatomy and motion conditioned over population characteristics and electrophysiology would enable several downstream tasks, including but not limited to personalised medicine and in silico trials, as well as contribute to a greater knowledge of the relationship among cardiac anatomy, motion and pathologies.

The proposed project would contribute a novel multi-modal geometric deep-learning model for the human heart, which would integrate cardiac geometry with motion and explore their relationship with population demography and electrophysiology. This, in turn, would contribute to the design of in silico trials, which have the potential to improve the efficiency of clinical trials and reduce costs. In silico trials also offer the possibility of simulating the efficacy of interventions in sub-populations underrepresented in the data, which would reduce spurious extrapolations of trial results from other populations to these groups. The model would also contribute towards the field of personalised medicine; by simulating the impact of interventions on a bespoke simulation of a subject's anatomy, personalised risk profiles for different interventions can be evaluated and compared. Finally, investigations into the relationships between cardiac anatomy, motion, demography, electrophysiology, and pathology could yield new insights into the aetiology of cardiovascular diseases.

The project fits within both Challenges 1 and 2 of the EPSRC Healthcare Technologies Research Theme strategy. For Challenge 1, the proposed approach will interpret and analyse population to understand both individual and population scale variation in disease phenotypes. The approach would also contribute towards the development of novel prediction tools like digital twins and the discovery of new indicators of susceptibility and risk of disease. For Challenge 2, the project would contribute to the development of novel techniques for optimising patient-specific illness prediction, biomarker identification, decision-support systems, and predicting susceptibility to illness.