AI Network Can Tell Age, Gender From Standard ECG

By Will Boggs MD

August 30, 2019

NEW YORK (Reuters Health) - A convolutional neural network trained through deep learning can accurately predict a person's age and gender using only standard 12-lead ECG signals, researchers report.

"Our standard diagnostic tools may have far more information behind them than we've come to expect throughout standard approaches to diagnostic interpretation," said Dr. Suraj Kapa from Mayo Clinic College of Medicine, in Rochester, Minnesota.

"Between this study and other prior studies showing that we can predict likelihood of having atrial fibrillation from a normal sinus ECG or the presence of a low ejection fraction, AI-enabled ECG analysis may offer new, rapid, and cost-effective insights into human health well beyond what we could have anticipated in the last two centuries since the ECG was first developed," he told Reuters Health by email.

Dr. Kapa's team investigated whether the application of AI algorithms to a large ECG patient data set could predict age and gender independent of additional clinical data and sought to determine whether discrepancies between ECG age and chronological age might be a marker of physiological health.

They trained the convolutional neural network (CNN) on nearly 400,000 patients' ECGs, internally validated it on nearly 100,000 ECGs, and further tested it on more than 275,000 additional ECGs.

The overall accuracy was 90.4% for females and 90.3% for males, with better accuracy for gender identification among people under 45 (93%) than among those older than 55 (89%).

As reported in Circulation: Arrhythmia and Electrophysiology, online August 27, the mean absolute error for estimating age was 6.9 years, and the accuracy for detection of age 40 years and older was 87%.

The CNN classified individuals into one of four age groups (18-25, 25-50, 50-75, and 75 and above) with 71.6% overall accuracy.

Of 100 patients with longitudinal data, those with close correspondence between their CNN-predicted age and chronological age (R-squared value of 0.9-1.0) had no significant medical comorbidities and were otherwise healthy.

In contrast, patients whose CNN-predicted age deviated from chronological age by more than seven years had a higher prevalence of pre-existing comorbidities (prior myocardial infarction, low ejection fraction, coronary disease, hypertension, and atrial fibrillation). After one of these patients underwent a heart transplant, the CNN-predicted age moved closer to that of the donor age.

"We are doing further research to understand how discrepancies between the AI predicted age and chronologic age might predict short- and long-term outcomes (mortality, incident major health events, etc.)," Dr. Kapa said. "These results may define utility of these AI-enabled predictive algorithms to create a whole new area of investigation in medicine that is truly scalable in a cost-effective manner to all patients."

"Implementation of these new tools (neural networks, machine learning, etc.) to medical data holds the promise not to replace physicians but to help clarify and augment insights into the health of the people we take care of," he added.

Dr. Peter W. Macfarlane of the Institute of Health and Wellbeing at the University of Glasgow, Scotland, wrote about the use of neural networks in ECG interpretation more than 25 years ago and recently reviewed the influence of age and sex on the ECG. He told Reuters Health by email, "At the present time, with all respect, this paper will not be of assistance to the practicing clinician. Who knows whether in 10 years an ECG-predicted age versus an actual age will have any meaning whatsoever."

"I do think that journal editors and physicians should not get carried away by the term AI quite yet," he said. "There will undoubtedly be a role for its use in medical diagnosis, but some degree of selection of the optimum use will be necessary."


Circ Arrhythm Electrophysiol 2019.