Images courtesy of CarDS Lab website.
According to the CDC, one person dies every thirty-four seconds from cardiovascular disease in the United States. It is also the leading cause of death for men and women across the country, costing over $200 billion annually. In 2020, around 697,000 people in the United States died from cardiovascular disease, which accounted for twenty percent of all deaths that year. There are strong efforts worldwide in research and clinical care to improve the diagnosis and treatment of this disease, particularly at the Cardiovascular Data Science (CarDS) Lab at the Yale School of Medicine, which is tackling this issue through a creative intersection of computer science and patient data.
The CarDS Lab aims to improve cardiovascular health using data-driven insights into how care is delivered to patients. This means they use technology—artificial intelligence (AI) with machine learning, for example—to augment our ability to diagnose and treat patients. For example, the lab has built AI models that can detect cardiac muscle dysfunction from electrocardiograms (EKG), which humans are unable to do. “The entire idea is to democratize the access to technology so that more people know they may have [cardiovascular disease] so that they can be referred to the health system for more advanced imaging,” said Yale University Assistant Professor of Cardiovascular Medicine Rohan Khera, who is the principal investigator of the lab. The group also works with national registries and datasets to define best methodological practices in conducting studies, evaluates healthcare policies and their association with cardiovascular health outcomes, and interprets clinical trial results personalized for individual patients. These goals are reflected in the structure of the lab, where members are part of a “core” or a specific aim within the lab’s overall goal. “Some folks work on natural language processing, some work on ECGs, some work on cardiac imaging, some are focused on EHR design, some are working on trials. People present from one theme to the others, so everybody can learn what the others are doing, but [they] tend to focus on their own domain. That’s been our key, to focus on building micro-labs within a large lab,” Khera said.
Khera grew up in India, where he attended the All-India Institute of Medical Sciences for his medical training. He then had a variety of away rotations at several institutions, including Johns Hopkins University, the University of California, Los Angeles, and the University of Pennsylvania, where he gained broad exposure to basic translational and clinical research. This experience continued at the University of Iowa during his residency and at UT Southwestern for his fellowship. “When I graduated fellowship, I knew I was going to start a research program…It felt like there was a lot happening at [Yale]. It was very exciting how [Yale] had been at the cutting edge of health policy and outcomes research, so I came here to see if I could extend that further into more advanced data science,” Khera said.
Khera started his faculty position at Yale in July 2020, during the peak of the COVID-19 pandemic. “Everything was shut down, and there was a lot of time spent thinking how one would structure the lab when nobody’s around,” Khera said. He noted that there were fewer opportunities to meet new people who would be interested in the lab, so he spent the first several months of the lab’s inception exploring the Yale community. He credits the decision to run the lab virtually as one of the key reasons for its success, as people do not need to be in the same room all the time. Now, the CarDS lab has grown to over twenty people.
Along this mission to integrate machine learning with cardiovascular health, the lab recently published a paper titled “Individualising intensive systolic blood pressure reduction in hypertension using computational trial phenomaps and machine learning: a post-hoc analysis of randomised clinical trials.” This study looked at two clinical trials, SPRINT and ACCORD BP, which each compared intensive versus standard blood pressure control treatment. Using a machine learning algorithm and a “phenomapping strategy,” which creates a network of all patients recruited in the trial to compare their phenotypes, they found that not every patient in the SPRINT trial benefitted to the same extent from the intensive blood pressure control treatment. In other words, the effect of the treatment seemed to vary across different types of patients. From these results, the lab was able to analyze a given patient’s key characteristics and can tell how likely that patient is to benefit from intensive blood pressure control treatment. They then validated these findings independently in the ACCORD BP trial. “The traditional interpretation of clinical trials does not necessarily inform us about whether a given treatment works for each patient…We’re interested in better understanding how the results of a study can be individualized for each unique patient in front of us at the clinic…We think that’s pretty interesting because not every patient should be treated in the same way. And that’s one step closer to more personalized cardiovascular care,” said Evangelos Oikonomou, a clinical fellow in cardiovascular medicine at Yale
A second project that the lab has been working on is developing AI models to diagnose structural heart diseases from printed ECG scans. The focus on ECG comes from the fact that they are the most widely accessible and ubiquitous tool in the world to better understand a patient’s heart. However, physicians are only able to diagnose certain conditions and heart rhythm disorders from ECGs, creating the need for more expensive and harder-to-obtain screening tools for other heart conditions. The goal of the project is to be able to diagnose these conditions from ECGs—leading this effort is Yale College senior Veer Sangha YC ’23, who has recently received the Rhodes Scholarship for his work with the lab. “We have a large repository of patients at the hospital, so we have their ECGs, and we know which patient has which disease,” Sangha said.“So we can train our deep learning models to be able to learn features in the ECG that are relevant to a certain class of disorders or a certain disorder that a patient may have. And it can learn these features that humans themselves cannot learn.” To make this further accessible for patients, Sangha developed the model so that it didn’t need to use the signal data from the ECGs, which is not always available at the point of care. Instead, the model can make these inferences from printed scans of an ECG, which are widely available to patients and their clinicians.
For undergraduate students interested in joining the CarDS lab, Khera recommends using the lab’s website Contact page or reaching out to him directly. He also suggests that students who want to join should ideally be interested in health technology and its applications and have some coding experience. Finally, he enjoys having students who want to be part of the lab for a long period of time. “Those who engage for a longer time are always there in a community learning, adapting, and growing. The folks that have really developed their careers in the lab have been associated with us for a couple of years already now,” Khera said.
You can learn more about the CarDS Lab at https://www.cards-lab.org/.