Novel AI algorithm detects liver disease via cardiovascular screenings
Key takeaways:
- Researchers envision training the algorithm to help cardiologists identify liver disease from subcostal views.
- A few next steps are needed before researchers can test the algorithm out in the clinic.
An AI-based algorithm may help detect chronic liver disease using images from standard echocardiography, according to study results published in NEJM AI.
“For both cardiologists and hepatologists, there is a growing understanding that these two fields may interact more than we expected in the past,” Alan C. Kwan, MD, senior author and assistant professor in the department of cardiology in the Smidt Heart Institute at Cedars-Sinai, told Healio.
Kwan and colleagues used more than 1.5 million echocardiogram videos from nearly 25,000 patients to develop and evaluate a deep-learning program designed to identify cirrhosis and steatotic liver disease. The researchers compared the AI predictions with diagnoses made from abdominal ultrasound or MRI studies.
Healio spoke with Kwan about the study, how the algorithm works and what additional research is needed before the technology can be utilized.
Healio: What prompted this study?
Kwan: I do research in cardiometabolic disease, and one of the focus areas for me has been how disease in the liver, normally steatotic or fatty liver disease, affects disease in the heart. There are all sorts of traditional risk factors — high blood pressure, cholesterol, insulin resistance, diabetes, obesity — but we have had difficulty organizing it all in a way that helps us understand metabolic health.
As a cardiologist, I have been thinking more about the liver as an “effector organ” on cardiovascular health. I read echocardiography studies, and I realized that while primarily use them to look at the heart, we also see multiple views of the liver. Specifically, in subcostal views, we look at the inferior vena cava, which gives us an idea about how much fluid is in the body, but on either side of the inferior vena cava, we can also see hepatic tissue. My thought was that maybe we could actually use this information instead of ignoring it.
I collaborate a lot with David Ouyang, MD, a researcher in AI and computer vision, and I figured that instead of trying to retrain cardiologists to identify liver disease, we could train an algorithm to help cardiologists identify liver disease from these subcostal views.
Healio: How does the algorithm work?
Kwan: The algorithm uses a transthoracic echocardiogram, which is comprised of more than 50 video images obtained from multiple different areas of the chest.
The algorithm takes the entire study and identifies images from the subcostal view, which includes images of the liver. The algorithm then picks out high-quality views of the liver and gives us a prediction of the presence of liver steatosis or cirrhosis.
Healio: What were the key study findings?
Kwan: Our algorithm did relatively well for identifying both steatosis and cirrhosis, though the ability to identify cirrhosis was better than the ability to identify steatosis, which may be because cirrhosis is more of an advanced disease state.
We developed this algorithm internally with our original datasets, but also had a dataset not used in training for validation, in which the algorithm remained relatively accurate. Additionally, we used echocardiograms from an external site (Stanford University) and compared the algorithm to other diagnostic modalities, like MRI, and it also performed relatively well in all those different settings.
Healio: Did any of the findings surprise you?
Kwan: Because ultrasound is so commonly used in hepatology to diagnose liver disease, there was some degree of fundamental biological plausibility that another ultrasound imaging modality could perform in a similar way. Thus, I thought this had a high probability of working and also thought it may be a good use of data that we are just throwing away.
Imaging studies provide a lot of information. As we process and read them, and create reports go out to the ordering physicians, there is a lot of information that is lost. One of the big pushes of both my research and Dr. Ouyang’s lab has been to try and use all the information we can get from these imaging studies, and this work is a good example of what we could potentially do.
Healio: What should our readers do with this information?
Kwan: Right now, this research is still in the early stages. There are a few next steps we need to do including testing in more clinical settings. We are planning to apply the algorithm prospectively in patients without known liver disease who have had echocardiograms performed, and then identify those with a high probability of having undiagnosed liver disease with our algorithm. We could then ask them to come in for confirmatory testing to see how our algorithm performs under those settings.
Metabolic dysfunction-associated steatotic liver disease (MASLD) affects approximately one in three US adults. The No. 1 cause of death among these patients is not primary hepatic disease although MASLD can lead to cirrhosis and liver cancer but is cardiovascular disease. Thus, when we think holistically about cardiovascular disease, considering non-traditional risk factors like MASLD may be increasingly important. There are now drugs that can actually treat forms of steatotic liver disease, which was not the case 2 or 3 years ago. In fact, we are seeing more and more drugs which may improve metabolic health, and appear to have off-target cardiovascular benefits, and thus there is a major opportunity here which needs further research.
Healio: What else would you like to emphasize?
Kwan: AI in medicine is a hot topic. However, it is important that we specify what we mean by AI to better focus the impact on how we care for patients. Originally, AI in medicine would make “black-box” type predictions, linking unrelated inputs and labels, which was novel, but not very clinically helpful. This has changed quickly, and there are now multiple directions that we are moving in, from improving image processing, to providing clinical decision support, to aiding complex data analysis.
In particular, in our research we focus on not just on the novelty of these AI algorithms to predict things but also to try and make connections that leverage all the information available in diagnostic images using computer vision. This approach has potential for significant cost savings. If we can use diagnostic images we already acquired for other reasons and get more information out of them, that is likely a pretty good benefit to the overall system.
However, this also highlights why understanding clinical context and testing this algorithm in real-world settings is so important. If our algorithm identifies disease that causes us to do more downstream testing, we need to make sure that this has a positive net impact vs. just increasing costs for something we cannot address.
There are a lot of complexities in AI and imaging and thinking broadly about it — more than just the novelty of it — is important, as this field is moving very quickly.
For more information:
Alan C. Kwan, MD, can be reached at alan.kwan@cshs.org.
Collapse