Individualized prediction of progression to heart failure using digital twins, an explainable reinforcement learning approach

The National Heart, Lung, and Blood Institute awarded the University of Illinois $109,076 on September 30, 2025, under the Cardiovascular Diseases Research program (CFDA 93.837) to develop individualized prediction methods for heart failure progression using digital twins and explainable reinforcement learning.

The project will characterize cardiac performance at the individual level using primary echocardiographic images, quantitative echocardiographic measures, and arterial pulse wave velocity data from 4,483 participants in the longitudinal NHLBI-funded Atherosclerosis Risk in Communities (ARIC) study who are free of heart failure. Researchers will apply domain-guided mechanistic reinforcement learning to simulate cardiac adaptation and detect how individual comorbidity profiles differentially modify pathophysiologic mechanisms of heart failure through a virtual digital heart model. The work addresses the gap between population-level heart failure risk identification and individualized prediction of whether and when a given person will develop symptomatic heart failure, accounting for the heterogeneity in heart failure susceptibility driven by extra-cardiac comorbidities.

The award is structured as a Project Grant, assistance type for competitive research awards. Performance occurs in Illinois and runs through September 29, 2030.

Abstract

Heart failure (HF) will affect one in four americans during their lifetime. IT is associated with a 50% 5-year mortality, disproportionately burdening individuals who experience accelerated decline in cardiac function. In most patients, cardiac dysfunction starts with an adaptive asymptomatic stage, before eventually progressing to symptomatic HF. Although current methods are excellent to identify population level HF risk, they are limited in identifying if or when any given individual develops HF symptoms. Extra-cardiac comorbidities are likely important contributors to heterogeneity observed in HF susceptibility and clinical presentation. Critically, traditional non-invasive measures of cardiac function perform poorly in discriminating HF patients from their comorbidity-matched symptom-free counterparts. The lack of effective approaches to identify individuals at risk of HF is a critical barrier to identifying and preventing HF. There is, thus, an urgent need to develop personalized methods to understand the natural history and progression to overt HF on an individual level. Our goal is to study the effect of an individual's comorbidity profile on their cardiac structure and function in early HF using explainable machine learning. Our central premise is that individual comorbidity profiles differentially modify pathophysiologic mechanisms of HF, which can be detected using a virtual digital heart, a "digital twin". Leveraging rich data from 4,483 participants from the longitudinal nhlbi-funded atherosclerosis risk in communities (aric) study free of HF, we will characterize cardiac performance at an individual level using primary echocardiographic images, quantitative echocardiographic measures, and arterial pulse wave velocity. We will then use domain-guided mechanistic reinforcement learning to simulate cardiac adaptation, conditioning on known HF-related comorbidities including body composition, obesity, renal function, pulmonary function, skeletal muscle strength, and frailty. Additionally, we will use circulating serum protein biomarkers to evaluate biological mechanisms of HF. We will leverage the longitudinal nhlbi-funded atherosclerosis risk in communities (aric) study. Our specific aims are to (1) define individualized comorbidity-driven risk of HF in asymptomatic older adults; and (2) to identify echocardiographic patterns in asymptomatic older adults that predict future HF onset. These studies will yield a novel personalized model of HF progression, providing critical insights into cardiac dysfunction at an asymptomatic stage and its trajectory. This study will be performed by pranav dorbala, an MD/PHD student at the university of illinois urbana champaign pursuing a PHD in machine learning. His sponsor has extensive expertise in machine learning for health care applications and his CO-sponsor is an established aric investigator with extensive experience in cardiovascular epidemiology, quantitative echocardiography, and high dimensional data. Completion of the proposed studies will serve to train the student in the fundamentals of becoming a physician-scientist in the field of personalized medicine.