Exploring the evolution of AI in cardiology

He added that AI now plays a major role in automating time-consuming tasks such as image segmentation and quantification, and is seeing use in procedural guidance and even outcome prediction, particularly in structural heart disease and heart failure interventions. He said the use if AI has occurred on a much faster timeline that he expected a decade ago.

"I didn't know that the exponential trend of AI was going to hit so quickly, and now I think everybody's using AI," Sengupta said. 

Where AI excels and struggles

Sengupta emphasized the dual nature of AI’s role in cardiology; it has great promise and great limitations. Imaging applications using convolutional neural networks are already improving efficiency and accuracy in diagnosing conditions like coronary artery disease. AI can effortlessly process and quantify plaque characteristics from coronary CT angiography (CCTA), tasks that would be impractical for humans to perform manually at scale.

However, more complex syndromes, like heart failure or certain valvular heart diseases, reveal the limitations of current AI models.

"AI does pattern recognition ... but [AI models] don't interpret the world the way we interpret the world. The human mind is completely different," Sengupta said.

Sengupta warned against overestimating AI’s capabilities. AI models are not human minds and they do not interpret things the same way as humans. While deep learning models may mimic aspects of human cognition, they lack contextual awareness. This disconnect can introduce bias and lead to incorrect conclusions, particularly when models are applied in diverse clinical settings, he said.

He cited studies that demonstrate how biased AI outputs can mislead physicians and worsen diagnostic outcomes. Additionally, AI models often lack transparency in how they reach conclusions, further complicating their integration into practice.

Barriers to AI implementation

A key barrier to the broader adoption of AI in cardiology is variability in hospital infrastructure. Imaging systems, electronic medical records and data formats differ widely, making it difficult to generalize AI models across institutions.

"So one model that has been developed in one university may not work elsewhere. The generalizability may not be there," he said.

He also warned that feeding AI models too much information and too many variables can sometimes lead to mistakes.

"I think there are clear opportunities like in imaging, and then there are some more difficult areas where there's a lack of clarity on the definitions and consensus on even how we define heart failure or where we can use it. In heart valve diseases, for example, how do you define severe disease or moderate disease? It's not all a bunch of numbers; there are lots of different aspects of the disease. Sometimes a human mind goes and looking at the images to finally qualify the patient for having certain types severity of the disease. And that's another area which can be refined by the application of AI," he said.

AI is about augmenting intelligence, no replacing humans

Looking ahead, Sengupta sees AI as a tool to augment rather than replace physicians. One area of growing interest is risk stratification. Traditional tools like the ASCVD or Framingham scores may soon be supplanted by AI models capable of analyzing a patient’s full clinical history, genetics and imaging data to provide personalized risk profiles.

Another promising direction is redefining disease taxonomy. 

"Just like genomics has changed how we classify cancers, AI could help us redefine heart failure or aortic stenosis based on new data patterns,” he said.