Abstract 11190: Subgroup Comparison of Electrocardiogram Deep-Learning Model Performance for Estimating Coronary Artery Calcium Score

Abstract

Introduction: Despite the clinical usefulness of Coronary artery calcium score (CACS), it is difficult to use it as a routine practice due to its high cost and radiation hazard. Recently, several deep learning (DL) models to estimate CACS using 12-lead electrocardiography (ECG) have been introduced. However, their performance has not been compared among different patient groups. This study compared the performance of DL models to estimate CACS in different patient groups. Methods: The data of 5,427 patients who underwent both CT scan and 12-lead ECG at Korea University Anam Hospital from November 2012 to December 2021 was split into 80% training set and 20% test set. CNN-based DL model for predicting CACS ≥ 1 and CACS ≥ 100 using the raw ECG waveforms and several clinical factors as input was developed. Results: The DL model showed comparable performance for CACS estimation in the test dataset (area under the receiver operating characteristics curves [AUROCs] 0.77 for CACS ≥ 1 and 0.82 for CACS ≥ 100). The DL model showed better performance in patients with cardiovascular risks less than 3 compared to patients with cardiovascular risks more than 3 (AUROC 0.75 vs 0.71 for CACS ≥ 1, 0.77 vs 0.71 for CACS ≥ 100, Table 1). Among subgroups, the elderly (age > 65 year-old) showed the lowest performance (AUROC 0.72 for both CACS ≥ 1 and ≥ 100). Interestingly, the model performance predicting CACS ≥ 1 is better in patients with normal sinus rhythm, whereas the model performance predicting CACS ≥ 100 is better in patients with non-normal sinus rhythm. This suggests that normal ECG can better predict healthy condition and abnormal ECG can better predict pathological condition. Conclusions: The performance of CACS estimation DL model could depend on the patient's cardiovascular risk factors. When applying the DL model to actual clinical practice, it is necessary to consider the performance of the DL model for the patient group to be applied and to selectively apply the appropriate DL model.