TLDR: Researchers developed a probabilistic framework that combines data imputation with mechanistic modeling to simulate heart-brain interactions, even when patient datasets have missing cardiac information. Using UK Biobank data, the model accurately imputes missing cardiac features and estimates cardiovascular parameters, enabling the exploration of how brain conditions like white matter hyperintensities relate to heart function.
Understanding the intricate connections between the heart and the brain is crucial for unraveling the mysteries of neurological conditions like vascular dementia and Alzheimer’s disease. However, a significant challenge in this research area is the scarcity of comprehensive patient datasets that capture detailed information for both organs simultaneously. Neuroimaging studies, for instance, often provide only limited cardiac measurements, typically just systolic and diastolic blood pressures, which are insufficient for a thorough analysis of cardiovascular factors in brain disorders.
Traditional approaches to studying heart-brain relationships often fall into two categories: statistical association models or purely mechanistic models. While statistical models are easy to formulate, they offer limited insight into the complex physiological interplay. Mechanistic models, on the other hand, provide deeper understanding of biomechanical aspects but are often “ill-posed,” requiring extensive data to tune their parameters accurately.
To bridge this gap, researchers Jaume Banus, Maxime Sermesant, Oscar Camara, and Marco Lorenzi have introduced a novel probabilistic framework that combines data imputation with mechanistic modeling. This innovative approach aims to learn cardiovascular dynamics from brain imaging and clinical data, even when cardiac information is incomplete. The framework leverages large-scale datasets, such as those found in the UK Biobank, to infer missing cardiac data and simulate realistic heart function.
At its core, the framework consists of two main components: a variational framework for imputing missing cardiac information and a Gaussian Process emulator. The imputation model predicts unobserved cardiac features based on available cardiac and brain data. Simultaneously, the Gaussian Process emulator mimics the behavior of a “lumped cardiovascular model,” which is a simplified yet effective representation of the heart’s mechanics. This combination allows for a comprehensive, end-to-end generative model that can both fill in missing measurements and estimate crucial cardiovascular parameters.
The study utilized a subset of 3445 subjects from the UK Biobank, including T1 and T2 FLAIR MRI images, along with various brain and cardiac image-derived phenotypes (IDPs). Brain IDPs included total grey matter, white matter, and ventricles volumes, as well as white matter hyperintensities (WMHs) volume and number – indicators of brain damage. Cardiac IDPs comprised stroke volume, ejection fraction, and end-diastolic volume for the left ventricle, alongside blood pressure measurements (DBP, SBP). These data were used to train and validate the model.
Experimental results demonstrated the effectiveness of this new method. It significantly outperformed conventional imputation techniques like mean and median imputation in accurately estimating missing cardiac features. Furthermore, the emulator consistently showed low errors in predicting the parameters of the lumped cardiovascular model. This means the model can reliably estimate and simulate cardiac function even from datasets with minimal cardiovascular information, such as only systolic and diastolic blood pressures.
One of the most exciting aspects of this framework is its ability to simulate cardiovascular dynamics and explore hypothetical scenarios of heart-brain relationships. For instance, the simulations revealed meaningful connections: an increase in white matter hyperintensities volume or brain volume loss was associated with decreased stroke volume and ejection fraction, indicating reduced heart contractility. Interestingly, the number of WMHs showed different dynamics, suggesting heart-remodeling processes like changes in ventricle size and contractility. These findings align with existing clinical literature, offering new insights into controversial topics like the pathogenesis of WMHs.
Also Read:
- DualAlign: Crafting Realistic Synthetic Clinical Data for Alzheimer’s Research
- Rethinking Brain Disease Biomarkers: The Case for Multimodal and Longitudinal Data
In conclusion, this generative model provides a powerful tool for analyzing complex physiological relationships between the heart and brain, particularly in datasets with limited available features. It offers a data-driven way to infer mechanistic parameters and explore various heart-brain scenarios through simulation. Future work aims to transfer this knowledge to other datasets, incorporate deep learning architectures, and include more complex cardiovascular features to further enhance its capabilities. For more detailed information, you can refer to the full research paper: Joint data imputation and mechanistic modelling for simulating heart-brain interactions in incomplete datasets.
