Physics-Guided AI Towards Better Diagnosis on Heart Diseases
We are developing physics-guided machine learning techniques for cardiovascular systems, to obtain a personalized digital twin model for heart diseases.
Heart disease, characterized by changes in vascular, valvular, and ventricular systems, has been the leading cause of death in US since 1950. These diseases often happen acutely, making early detection essential. This interdisciplinary project aims to transform the clinical approach to early stage heart diseases, which is still dominant by traditional methods, through an AI-based approach. Our goal is to integrate data from wearable devices, mechanical measurements, and physics-based modeling, towards a personalized digital twin model for heart diseases. To this end, the project has three sub-tasks. First, a physically interpretable constitutive modeling approach will be developed, to capture the change of mechanical responses and the underlying mechanisms for cardiovascular tissue degradation. Second, an attention-mechanism-based approach will be designed, to infer the correlation between multiphysics data and the corresponding tissue model. Third, an efficient surrogate model will be constructed, enabling fast inference to assist in real-time monitoring. As the long-term goal, this personalized digital twin model is anticipated to provide early risk assessments, while also offering insights into cardiovascular mechanics that could lead to new treatment strategies.
In the past semester, our team has made progresses on the first and the second sub-tasks. For the first sub-task, a peridynamic neural operator model has been developed, which constructs the constitutive law as well as the underlying fiber orientation from loading-response measurements. To
demonstrate the applicability of our approach, we apply the HeteroPNO in learning a tissue model and fiber orientation field from DIC measurements of a heart valve leaflet specimen. For the second sub-task, we have developed an attention-mechanism-base approach to extract global prior information from mechanical measurements of multiple tissue specimens, and provide a generalizable foundation model to new and unseen tissues. As the ongoing work, we are extending this model to multiphysics and multimodal data.
