Skip to content

AI-based Analysis of ECGs Complements CHARGE-AF Score for Predicting AF

Key findings

  • Within a sample of 45,770 individuals receiving longitudinal primary care at Massachusetts General Hospital, researchers developed ECG-AI, a convolutional neural network that predicts time to incident atrial fibrillation (AF) using 12-lead ECG data
  • In three independent datasets, AF risk estimates by ECG-AI were comparable to those derived from the CHARGE-AF (Cohorts for Heart and Aging in Genomic Epidemiology–AF) score, which requires 11 pieces of information
  • CH-AI, a model that combines ECG-AI and CHARGE-AF, consistently outperformed CHARGE-AF alone within the three independent datasets, which included 83,162 individuals whose clinical characteristics varied substantially
  • This study supports the utility of deep learning models based on 12-lead ECG data for determining AF risk and provides the first evidence that predictive value is maintained for up to five years after an ECG is performed

Over the past few years, several research groups have developed artificial intelligence (AI)–based computer models that can predict atrial fibrillation (AF) from an electrocardiogram. These models don't explicitly estimate event-free survival, though, and it isn't clear whether they improve on the use of clinical risk factors.

Shaan Khurshid, MD, MPH, electrophysiology fellow, and Steven A. Lubitz, MD, MPH, cardiac electrophysiologist in the Telemachus & Irene Demoulas Family Foundation Center for Cardiac Arrhythmias and researcher in the Cardiovascular Research Center at Massachusetts General Hospital, and colleagues have trained a convolutional neural network (CNN) called ECG-AI to predict time to incident (newly diagnosed) AF/atrial flutter. Through exposure to tens of thousands of examples, a CNN "learns" to extract patterns from medical imaging data even when the appearance varies.

In Circulation, the team reports ECG-AI performed as well as the 11-component CHARGE-AF (Cohorts for Heart and Aging in Genomic Epidemiology–AF) score. A system comprising both CHARGE-AF and ECG-AI performed better than either component alone.

Study Methods

ECG-AI was developed using 100,954 single 12-lead ECGs from 45,770 individuals, ages 18–90, who received longitudinal primary care at Mass General between 2000 and 2019. It was trained to predict the probability of five-year AF-free survival.

The team then assessed ECG-AI in three independent datasets collectively involving 83,162 individuals whose clinical characteristics varied substantially: 4,166 additional individuals from Mass General, 37,963 from Brigham and Women's Hospital (BWH) and 41,033 from the UK Biobank.

The primary outcome was incident AF at five years. Due to limitations in available follow-up data, the primary outcome for the UK Biobank cohort was incident AF at two years.

Model Performance

The power to predict incident AF was expressed as the area under the receiver operating curve (AUC, maximum=1):


  • Mass General cohort: AUC, 0.82
  • BWH: 0.74
  • UK Biobank: 0.71


  • Mass General: 0.80
  • BWH: 0.75
  • UK Biobank: 0.73

CH-AI (combination of ECG-AI and CHARGE-AF)

  • Mass General: 0.84 (P<0.05 vs. CHARGE-AF)
  • BWH: 0.78 (P<0.05 vs. CHARGE-AF)
  • UK Biobank: 0.75 (P=0.28 vs. CHARGE-AF)

The improved performance with CH-AI versus CHARGE-AF was generally consistent among individuals with a history of heart failure or stroke, for whom AF risk assessment may be particularly important.

Clinical Outlook

This study suggests ECG-based risk signals derived from deep learning are as useful as clinical risk models for predicting incident AF. In addition, ECG-AI and clinical risk factors seem to provide complementary information that augments risk prediction.

The ability to predict incident AF up to five years in the future may encourage clinicians and patients to implement preventive measures such as alcohol cessation, weight control and treatment of hypertension. It should also increase the efficiency of AF screening by targeting individuals most likely to have AF identified with diagnostic testing.

In models developed using only a single ECG lead, CH-AI continued to perform better than CHARGE-AF. This intriguing finding suggests deep learning may facilitate AF risk estimation by wearable devices, which are commonly equipped with single-lead ECG capability.

Learn more about the Telemachus & Irene Demoulas Family Foundation Center for Cardiac Arrhythmias

Refer a patient to the Corrigan Minehan Heart Center


Using wrist-worn accelerometer data from the UK Biobank, Shaan Khurshid, MD, MPH, and Steven A. Lubitz, MD, MPH, of the Cardiovascular Research Center, and colleagues confirmed guideline-recommended levels of physical activity are associated with substantially lower risks of incident atrial fibrillation and stroke.


Shaan Khurshid, MD, and Steven A. Lubitz, MD, MPH, of the Cardiovascular Research Center, and colleagues found that among 493 patients with acute ischemic stroke, only 37% received cardiac monitoring within six months as guidelines suggest.