LOS ANGELES (July 11, 2024) -- The data fed to artificial intelligence (AI) systems make all the difference on performance, according to David Ouyang, MD, a cardiologist in the Department of Cardiology in the Smidt Heart Institute at Cedars-Sinai.
Ouyang, who is also a faculty member in the Division of Artificial Intelligence in the Department of Medicine, has authored several papers describing AI systems for analyzing datasets related to cardiovascular disease. He is the corresponding author of two new papers about the details of developing such systems.
One paper, a study published in JACC: Advances, investigates the selection of cases for an AI model to learn the impact on model performance. In this example, the models were trained to detect a cardiovascular condition called cardiac amyloidosis. This condition occurs when the body creates abnormal proteins that accumulate in the tissues of the heart, causing the heart to become stiff. This training was based on different definitions of the disease, and the study showed that this impacted the accuracy of the model.
“Because cardiac amyloidosis is underdiagnosed, an AI program trained on patients at high risk for the condition could help clinicians detect the condition,” Ouyang said.
Cedars-Sinai investigator Lily Stern, MD, also worked on the JACC: Advances study.
In another paper, a case study published in NEJM AI, Ouyang and colleagues used electrocardiogram (ECG) data to train deep-learning AI models to identify patients with heart failure, a condition in which the heart’s ability to pump blood is weakened. One program was asked to predict which patients corresponded with a diagnosis of heart failure in the patient record. The other was trained to predict the underlying measurement of disease severity.
The investigators found that it was more informative for the model to be trained on the clinical measurement used to diagnose heart failure, called ejection fraction, rather than a heart failure diagnosis in the electronic health record.
“This paper shows you will get more precise outcomes if you train the AI model on more granular detailed data, such as numerical measurements, rather than a description or diagnosis, which are less reliable,” said Amey Vrudhula, a fellow at Cedars-Sinai and first author of the study.
“We've trained a lot of AI models in healthcare and have learned important lessons in their design that we hope other researchers can use for future work,” Ouyang said.
Cedars-Sinai investigator Neal Yuan, MD, also worked on the study.
Read more in Cedars-Sinai Discoveries: The Doctor Is Still the Boss in the AI Era
RSS