Artificial Intelligence for Heart Function Assessment

Osaka Metropolitan University scientists have unveiled an innovative use of AI that classifies cardiac functions and pinpoints valvular heart disease with unprecedented accuracy, demonstrating continued progress in merging the fields of medicine and technology to advance patient care.

Chest radiographs, or chest X-Rays, are performed in many hospitals and very little time is required to conduct them, making them highly accessible and reproducible. Accordingly, the research team led by Dr. Daiju Ueda, from the Department of Diagnostic and Interventional Radiology at the Graduate School of Medicine of Osaka Metropolitan University, reckoned that if cardiac function and disease could be determined from chest radiographs, this test could serve as a supplement to echocardiography.

Dr. Ueda’s team successfully developed a model that utilizes AI to accurately classify cardiac functions and valvular heart diseases from chest radiographs. Since AI trained on a single dataset faces potential bias, leading to low accuracy, the team aimed for multi-institutional data. Accordingly, a total of 22,551 chest radiographs associated with 22,551 echocardiograms were collected from 16,946 patients at four facilities between 2013 and 2021. With the chest radiographs set as input data and the echocardiograms set as output data, the AI model was trained to learn features connecting both datasets.

The AI model was able to categorize precisely six selected types of valvular heart disease, with the Area Under the Curve, or AUC, ranging from 0.83 to 0.92. (AUC is a rating index that indicates the capability of an AI model and uses a value range from 0 to 1, with the closer to 1, the better.) The AUC was 0.92 at a 40% cut-off for detecting left ventricular ejection fraction—an important measure for monitoring cardiac function.

Reference: Artificial Intelligence-based Model to Classify Cardiac Functions from Chest Radiographs: Multi-institutional Model Development and Validation Study, The Lancet Digital Health, DOI 10.1016/S2589-7500(23)00107-3