Artificial Intelligence

Agentic AI for accelerated discovery
We are interested in understanding whether agentic AI can accelerate biomarker discovery in cardiovascular disease and help generate useful, testable hypotheses. In a recent collaboration with FutureHouse/Edison, we used the Kosmos platform to automate large portions of multi-omic causal inference and prioritization that previously required extensive manual bioinformatics. This agentic workflow rapidly highlighted miR-222-3p as a candidate causal regulatory node in fibrosis biology, motivating follow-up analyses and experimental validation to assess mechanism.

Agentic AI for multimodal cardiac interpretation
We build agentic, multimodal AI systems that interpret ECG, echocardiography, and cardiac MRI in ways that reflect how cardiologists reason across tests. By coordinating modality-specific expert models and synthesizing evidence into a unified assessment, our approach enables more accurate answers to complex clinical questions. In tasks requiring cross-modality synthesis, we see large gains over prior state-of-the-art models (>40 percentage points in diagnostic accuracy on multimodal cases).

We employ deep learning to design next-generation gene therapy cassettes (hypercompact CRISPR Cas systems) and vectors (adeno-associated virus and lipid nanoparticles) with the goal of creating new tools that we can use to not just treat, but potentially cure, inherited cardiovascular diseases. Our goal is to develop general methodology for AI powered in silico and multi-tissue ex vivo selection platforms for therapeutic development.

We build multimodal models to analyze raw cardiac signals such as Cardiac MRIs, Echocardiograms and ECGs.

We investigate how frontier AI models can enhance clinical cardiac care by democratizing specialized cardiovascular expertise, enabling healthcare providers across diverse settings to access sophisticated diagnostic and treatment guidance that was previously limited to major cardiac centers (Figure 1 and 2).

We develop and apply splicing models to decode complex RNA splicing patterns in cardiac tissues, providing insights into how genetic variants influence heart disease development and progression.

Our work in cardiac imaging extends to magnetic resonance imaging (cardiac MRI). We have developed an open-source software tool, DeepFlow, which uses a U-Net based segmentation model to extract aortic blood flow parameters, ranging from blood flow velocities to valve function. The deployment of DeepFlow on data from the UK Biobank, encompassing over 35,000 participants, has enabled insight into the genetic underpinnings of these cardiovascular traits. Causal inference through Mendelian randomization revealed that genes affecting aortic size may cause aortic valve insufficiency. These insights could lead to genetically-informed treatments.

We develop deep learning models to understand medical signals. We hope to build systems which improve physicians’ performance on everyday tasks like electrocardiogram interpretation, to discover new diagnostic abilities like distinguishing between hypertension and hypertrophic cardiomyopathy as causes of cardiac hypertrophy, and to develop new risk scores for understanding patients’ overall health like our SEER model. We also seek to better understand existing models, for example by investigating how they perform in different demographic groups and by trying to find the underlying features on which these techniques rely. Our group also develops methods for interpreting cardiac imaging, targeting well-understood as well as novel applications.

We use digital data to develop novel deep learning architectures for the detection of abnormal cardiac events. Monitoring and tracking of patients’ cardiovascular states are of particular interest and we have focused on two data streams in particular: wearable devices and in-hospital telemetry. We developed Deepbeat a novel deep learning method for the detection of atrial fibrillation from wearable device data. We are actively developing approaches to smart telemetry to improve in-hospital alarms.

The My Heart Counts Cardiovascular Health Study is one of the largest fully digital platforms to study how lifestyle patterns influence cardiovascular health. Our app was the platform for the first fully digital randomized clinical trial. Recently, we showed that personalized e-coaching prompts were more likely to increase physical activity. The next generation of the My Heart Counts Cardiovascular Health Study will leverage advances in artificial intelligence to deliver real-time insights and recommendations, tailored to a user’s unique physiology, routines, and situation, all while studying the impacts of physical activity on cardiovascular health and wellness as a whole.