Deep-learning algorithms applied to a widely available and low-cost test, the electrocardiogram (ECG), can improve the identification and classification of cardiac arrhythmias and the detection of early-stage heart disease, according to two studies published in Nature Medicine. “Automating arrhythmia detection can make heart monitoring with ECG more accessible and useful as a first-line diagnostic tool,” says Awni Hannun, corresponding author of one of the articles. “These results have the potential to lead to reduced rates of currently misdiagnosed, computerized ECG interpretations, and to improvements in efficiency of expert-human ECG interpretation.”

Credit: Getty/alengo

Hannun and colleagues used a large training dataset including 91,232 single-lead ECG records from 53,549 patients to develop a deep neural network to classify 10 arrhythmias, sinus rhythm and noise for a total of 12 output rhythm classes. “In the past, most work has focused on only a few of the sometimes easier to diagnose arrhythmias such as atrial fibrillation, but in this work we diagnose a much larger set including various forms of heart block,” explains Hannun. The researchers then validated the deep neural network on an independent test dataset (328 ECG records from 328 patients) that had been annotated by a consensus committee of board-certified cardiologists. Each ECG record in the test dataset was also classified by six cardiologists who were not part of the committee. The deep neural network matched the performance of board-certified cardiologists — the algorithm had an average area under the receiver operating characteristic curve (AUC), in which 1.00 would be a perfect classification and 0.50 a random classification, of 0.97 — and exceeded the average cardiologist sensitivity for all rhythm classes, with a higher average F1 score (the harmonic mean of the positive predictive value and sensitivity) than cardiologists (0.837 versus 0.780).

In another study, Attia et al. used artificial intelligence applied to the ECG for the detection of asymptomatic left ventricular (LV) dysfunction, a precursor of heart failure. The team used paired 12-lead ECG and echocardiogram data from 44,959 patients to train a deep-learning approach, known as a convolutional neural network, to identify patients with LV dysfunction (ejection fraction <35%) using only the ECG record. The network model was then validated on an independent dataset of 52,870 patients. The model had an AUC of 0.93 for diagnosis of mechanical dysfunction, and a sensitivity, specificity, accuracy and F1 score of 86.3%, 85.7%, 85.7% and 49.5%, respectively. Of note, 1,335 patients without LV dysfunction were assigned by the algorithm as having low ejection fraction. Attia et al. suggest that these ‘false positives’ are an early detection of ECG abnormalities, given that the risk of developing future LV dysfunction was higher in patients with an initial ‘false positive’ diagnosis than in those with a negative diagnosis (HR 4.1, 95% CI 3.3–5.0). The 5-year incidence of LV dysfunction in these patients was 9.5%.

These studies demonstrate the potential of deep-learning approaches to improve the accuracy and efficacy of ECG reading for patient screening and stratification.