Comparison of artificial intelligence and cardiologists in the interpretation of exercise treadmill tests

Abstract Introduction The exercise treadmill test (TMT) is commonly used as a screening test for ischemic heart disease. The sensitivity of TMT is approximately 50%, with a specificity of 90%. However, challenges in interpreting TMT results persist, even among cardiologists. Aim We aimed to develop an artificial intelligence (AI) model to interpret TMT results. The primary endpoint was the accuracy comparison between the AI model and cardiologists for screening obstructive coronary artery disease. Methods Out of a total of 800 TMTs performed on adults between 2018 and 2020, we used 500 for training, 200 for validation, and 100 as a test set. The deep learning model was based on convolutional neural networks (CNNs) developed using TensorFlow (Google, Mountain View, CA, USA) and Keras packages in Python. After training, the AI algorithm and an EP cardiologist independently interpreted 100 TMTs (test set). This assessment was subsequently reviewed by a blinded radiologist, who provided a final report of obstructive coronary artery disease (CAD, defined as ≥50% stenosis in at least one vessel) based on coronary computed tomography. Results In the test set of TMTs, the average patient age was 58.1 years (range: 20-78), with 68% male and 33% having obstructive CAD. The AI model's performance was compared to conventional TMT interpretation. The accuracy of the AI model was comparable to that of the cardiologist (0.720 vs. 0.680, p=0.845). The F1 score was slightly higher for the AI model than for the cardiologist (0.440 vs. 0.384) in TMT interpretation. Conclusions The AI model demonstrated performance comparable to that of a cardiologist in the interpretation of exercise treadmill tests. This suggests that AI has the potential to be a valuable tool in assisting with the analysis of exercise ECG data.