Electrocardiogram-Based Heart Age Estimation by a Deep Learning Model Provides More Information on the Incidence of Cardiovascular Disorders

Chiao-Hsiang Chang; Chin-Sheng Lin; Yu-Sheng Luo; Yung-Tsai Lee; Chin Lin

doi:10.3389/fcvm.2022.754909

Electrocardiogram-Based Heart Age Estimation by a Deep Learning Model Provides More Information on the Incidence of Cardiovascular Disorders

Front Cardiovasc Med. 2022 Feb 8:9:754909. doi: 10.3389/fcvm.2022.754909. eCollection 2022.

Authors

Chiao-Hsiang Chang¹, Chin-Sheng Lin¹, Yu-Sheng Luo², Yung-Tsai Lee³, Chin Lin^{2

4

5}

Affiliations

¹ Division of Cardiology, Department of Internal Medicine, Tri-Service General Hospital, National Defense Medical Center, Taipei, Taiwan.
² Graduate Institute of Life Sciences, National Defense Medical Center, Taipei, Taiwan.
³ Division of Cardiovascular Surgery, Cheng Hsin Rehabilitation and Medical Center, Taipei, Taiwan.
⁴ School of Medicine, National Defense Medical Center, Taipei, Taiwan.
⁵ School of Public Health, National Defense Medical Center, Taipei, Taiwan.

Abstract

Objective: The biological age progression of the heart varies from person to person. We developed a deep learning model (DLM) to predict the biological age via ECG to explore its contribution to future cardiovascular diseases (CVDs).

Methods: There were 71,741 cases ranging from 20 to 80 years old recruited from the health examination center. The development set used 32,707 cases to train the DLM for estimating the ECG-age, and 8,295 cases were used as the tuning set. The validation set included 30,469 ECGs to follow the outcomes, including all-cause mortality, cardiovascular-cause mortality, heart failure (HF), diabetes mellitus (DM), chronic kidney disease (CKD), acute myocardial infarction (AMI), stroke (STK), coronary artery disease (CAD), atrial fibrillation (AF), and hypertension (HTN). Two independent external validation sets (SaMi-Trop and CODE15) were also used to validate our DLM.

Results: The mean absolute errors of chronologic age and ECG-age was 6.899 years (r = 0.822). The higher difference between ECG-age and chronological age was related to more comorbidities and abnormal ECG rhythm. The cases with the difference of more than 7 years had higher risk on the all-cause mortality [hazard ratio (HR): 1.61, 95% CI: 1.23-2.12], CV-cause mortality (HR: 3.49, 95% CI: 1.74-7.01), HF (HR: 2.79, 95% CI: 2.25-3.45), DM (HR: 1.70, 95% CI: 1.53-1.89), CKD (HR: 1.67, 95% CI: 1.41-1.97), AMI (HR: 1.76, 95% CI: 1.20-2.57), STK (HR: 1.65, 95% CI: 1.42-1.92), CAD (HR: 1.24, 95% CI: 1.12-1.37), AF (HR: 2.38, 95% CI: 1.86-3.04), and HTN (HR: 1.67, 95% CI: 1.51-1.85). The external validation sets also validated that an ECG-age >7 years compare to chronologic age had 3.16-fold risk (95% CI: 1.72-5.78) and 1.59-fold risk (95% CI: 1.45-1.74) on all-cause mortality in SaMi-Trop and CODE15 cohorts. The ECG-age significantly contributed additional information on heart failure, stroke, coronary artery disease, and atrial fibrillation predictions after considering all the known risk factors.

Conclusions: The ECG-age estimated via DLM provides additional information for CVD incidence. Older ECG-age is correlated with not only on mortality but also on other CVDs compared with chronological age.

Keywords: artificial intelligence; cardiovascular diseases; chronological age; deep learning; electrocardiography.