Deep Learning Predicts Not Only Coronary Artery Disease Diagnosis But Also Prognosis

Published in the International Journal 'Radiology: Artificial Intelligence'

A deep learning model has been developed that can not only diagnose coronary artery disease using vascular CT scan results but also predict the future risk of heart disease.

Jin Heo, Professor of Radiology at Severance Hospital. Severance Hospital

The research team led by Jin Heo, Professor of Radiology at Severance Hospital, announced on the 19th that, together with Jinyoung Kim, Professor of Radiology at Keimyung University Dongsan Hospital, and the medical imaging AI company Pantomics, they have developed a deep learning model for diagnosing and predicting the prognosis of coronary artery disease and have verified its potential for real-world clinical application.

For patients visiting the emergency room with acute chest pain, it is crucial to quickly and accurately diagnose coronary artery disease and assess the risk of heart disease. Although CT angiography is performed to diagnose and assess these risks, there are limitations such as the lengthy time required for result interpretation and the possibility of varying interpretations depending on the reader.

The research team utilized artificial intelligence deep learning technology to automatically interpret coronary artery stenosis and developed a model that classifies patients into three groups?normal, non-occlusive (less than 50% stenosis), and occlusive (50% or more stenosis)?based on the degree of stenosis, and evaluated its accuracy.

The deep learning model was trained on data from 408 patients who visited the emergency rooms of three university hospitals and underwent CT angiography between 2018 and 2022. In addition, the team used the YOLO architecture to increase the speed of detecting vascular stenosis. The YOLO architecture offers the advantage of fast data processing because it performs object localization and classification simultaneously.

To validate the effectiveness of the deep learning model, the occurrence of cardiac events in all patients was tracked for an average of 2 years and 6 months. Among the patients, 15% experienced hospitalization or death due to conditions such as myocardial infarction or unstable angina. In particular, the incidence rate in the occlusive group was 38.8%, which was significantly higher than in the normal group (0.6%) and the non-occlusive group (3.2%).

The study also found that the degree of occlusion analyzed by deep learning was the most effective indicator for predicting future heart disease risk. When deep learning was used to add coronary artery occlusion to existing risk factors for risk analysis, the discriminative power for predicting future heart disease increased by 14% compared to using only existing risk factors (discriminative power of 80%).

Professor Jin Heo stated, "This study suggests that deep learning models can be applied in emergency rooms, where rapid diagnosis and treatment decisions are critical, not only to determine the presence or absence of coronary artery disease but also to predict patient prognosis," adding, "We have confirmed that artificial intelligence technology can expand beyond simple diagnostic assistance to become a clinical decision support tool."

Meanwhile, the results of this study were published in the international journal 'Radiology: Artificial Intelligence.'