Debra L. Beck and Eugene Braunwald, MD 

The use of computers to replicate human skills in medical diagnosis has been an active topic of research in recent years. In two papers published in Circulation, researchers presented findings from two such efforts.

Giudicessi et al sought to train and validate an artificial intelligence (AI)-enabled 12-lead ECG algorithm to determine the QTc, and then prospectively test this algorithm on tracings acquired from a mobile ECG (mECG) device in a population enriched for repolarization abnormalities. To do this they used >1.6 million 12-lead ECGs from 538,200 patients and derived and validated a deep neural network (DNN) to predict the QTc using cardiologist-overread QTc values as the “gold standard.” The ability of this DNN to detect clinically-relevant QTc prolongation (eg, QTc ≥500 ms) was then tested prospectively on 686 patients with genetic heart disease (50% with long QT syndrome).

Strong agreement was observed between human over-read and DNN-predicted QTc values in the validation sample and the prospective, genetic heart disease-enriched dataset. When applied to mECG tracings, the DNN’s ability to detect a QTc value ≥500 ms yielded an area under the curve, sensitivity, and specificity of 0.97, 80.0%, and 94.4%, respectively, leading the investigators to conclude that an AI DNN can be used to predict accurately the QTc of a standard 12-lead ECG.

Raghunath et al designed a DNN to predict new-onset AF from the resting 12-lead ECG. They used >1.6 million resting 12-lead digital ECG traces from 430,000 patients without a history of AF to train and validate their DNN. The also sought to identify the patients at risk for AF-related stroke among those predicted to be high risk for AF by the model at different prediction thresholds.

For their DNN, the area under the receiver operating characteristic curve and area under the precision-recall curve were 0.85 and 0.22, respectively, for predicting new-onset AF within 1 year of an ECG. In a simulated deployment scenario, the model predicted new-onset AF at 1 year with a sensitivity of 69% and specificity of 81%. The model also identified patients at high risk for new-onset AF in 62% of cases where an AF-related stroke occurred within 3 years of the index ECG. 

Comments 

In an editorial comment, Dr. M Rosenberg noted with interest that these two studies take “nearly opposite” approaches to the use of machine learning for ECG interpretation. The first study attempts to use computers to predict a well-known ECG parameter better than humans, while the second study tries to predict “future risk agnostic to any previously described ECG parameters,” given there is no well-established prognostic ECG metric for the prediction of AF. Both have the potential to greatly improve health care delivery. However, before AI can be used for these and other purposes, “we must identify ways to interpret, understand, and trust the magic behind these incredibly accurate and efficient prediction algorithms,” he wrote and this process is still in its infancy.

Braunwald’s Heart Disease: A Textbook of Cardiovascular Medicine remains the most trusted reference in the field and the leading source of reliable cardiology information for practitioners and trainees worldwide. Written and edited by global experts in the field, this award-winning text is an unparalleled multimedia reference for every aspect of this complex and fast-changing area. Dr. Braunwald also curates the extensive, bimonthly online updates that include “Hot Off the Press,” “Practice Updates,” and “Late-Breaking Clinical Trials.”

Preorder Braunwald’s Heart Disease: A Textbook of Cardiovascular Medicine, 12th Edition here! Publishing November 2021.