I wrote this article about a year ago and it was originally published on the GE website.
My conclusion is that although AI ECG is promising, the technology needs further clinical validation from additional studies conducted on varying populations before it is ready to become a useful clinical adjunct to the practicing cardiologist.
In the last five years, there has been an explosion of research on applications of artificial intelligence in medicine. Computer programs utilizing neural networks and large data sets could offer promising assistance to doctors as they work to glean diagnoses and prognoses from healthcare data that otherwise might have been overlooked.
Because electrocardiography is inexpensive, noninvasive, and ubiquitous, early research has focused on this modality.
The norm for interpreting a standard twelve-lead ECG is an initial computer-generated interpretation followed by a cardiologist overread. However, it is well recognized, including by the American Journal of Medicine, that standard computer ECG interpretation can lead to substantial inaccuracies in diagnosing cardiac rhythm, left ventricular hypertrophy, and myocardial infarction.1
Fortunately, there is now an intriguing possibility that AI could:
- Improve the accuracy of human/computer ECG diagnoses of cardiac rhythm and hypertrophy
- Help predict systolic function and mortality
As ongoing research further explores these areas, it is clear that AI will play a larger role in cardiac decision-making and patient care.
Improved Accuracy of ECG Rhythm Interpretation
Several studies in the last few years have examined AI that utilizes deep neural networks (DNN) to diagnose cardiac rhythm abnormalities.
These studies use a type of machine learning, or deep learning, that mimics the function of brain neurons using a series of interconnected computational statistical algorithms or nodes. Once the system is trained on a large data set of ECGs, the DNN performance is tested on an independent data set and assessed for its sensitivity, specificity, and overall accuracy.
A study published in Nature analyzed single-lead ECGs obtained from an FDA cleared, patch-based ambulatory ECG monitor.2 The DNN-based analysis demonstrated a superior ability to classify twelve rhythm classes compared to an individual “average” cardiologist interpretation using a consensus committee of board-certified cardiologists as the gold-standard. The positive predictive value and sensitivity for the DNN (0.837) was significantly higher than that of average cardiologists (0.780).
A study in the Journal of Electrocardiology evaluated a proprietary DNN-enhanced ECG algorithm for full twelve‑lead ECG analysis, including rhythm and ischemic abnormalities.3 The AI-enhanced analysis showed significantly higher accuracy (92.2% versus 87.2%) in identifying a major ECG abnormality compared to the conventional algorithm with similar sensitivity but improved specificity (94.0% versus 84.7%).
A follow-up study from the same group published in IJC Heart and Vasculature focused on the diagnosis of Afib.4 The DNN algorithm again proved significantly more accurate than the conventional computer algorithm (91.2% versus 80.2%) and equal to the physician overread plus the conventional algorithm.
As AI advances ECG interpretation for rhythm abnormalities, cardiologists will likely see it improve the utility of personal wearable ECGs and the efficiency and accuracy of long-term monitor ECG analysis.
Enhanced ECG Diagnosis of Hypertrophy
The twelve-lead ECG has long been utilized in assessments of cardiac structural changes, including enlargement of the atria and hypertrophy of the ventricles. However, the limitations of human interpretation, even with standard computer assistance, are well recognized.
A study in Europace showed that an AI algorithm applied to twelve-lead ECGs improved the accuracy of echocardiographic detection of left ventricular hypertrophy (LVH) over classic ECG LVH criteria and cardiologist interpretation.5 At the same specificity, the sensitivity of the AI algorithm was 159.9%, 177.7%, and 143.8% higher than those of the cardiologist’s assessment, the classic Sokolov-Lyon criteria, and the standard computer ECG diagnosis.
Another study, published in the Journal of the American College of Cardiology, found that a DNN was able to identify hypertrophic cardiomyopathy better than standard ECG criteria for LVH.6 The model performed particularly well in younger patients, with 95% sensitivity and a specificity of 92%.
While these findings hold promise, exactly how cardiologists would incorporate AI-enhanced ECG LVH prediction into their workflow requires further investigation.
Predicting Afib and Low Ejection Fraction Using AI-enhanced ECG
Cardiologists have long accepted that the ECG provides information on the structure of the heart, yielding valuable but limited insights into atrial and ventricular hypertrophy. But a paper in Nature Medicine demonstrates that AI-enhanced ECG analysis can predict LV systolic function.7
Mayo Clinic researchers paired twelve-lead ECG with echocardiographic data from 44,959 asymptomatic individuals and trained a DNN to identify patients with left ventricular systolic dysfunction (LVSD) (defined as ejection </= 35%, determined from paired echocardiograms) using the ECG.
The DNN showed a sensitivity of 86.3%, a specificity of 85.5%, and an overall accuracy of 85.7% in diagnosing LVSD. Those patients without LVSD but with an abnormal DNN result faced a four-fold increased risk of developing future LVSD compared to those with a normal result.
The Mayo Clinic group also demonstrated in a Lancet article that their DNN model can identify subtle ECG differences that predict which patients have had Afib.8 The study found that, by analyzing a single twelve-lead ECG obtained when patients were in normal sinus rhythm, the DNN model predicted which patients had previously been identified with Afib with 79.0% sensitivity, 79.5% specificity, and 79.4% accuracy.
If these preliminary observations can be confirmed in other centers and populations, an AI-enhanced twelve-lead ECG analysis could serve as a trigger for additional testing in the form of long-term monitors or echocardiograms, with the goal of reducing stroke risk and cardiac mortality.
Evidence for Predicting Mortality with AI
If AI-enhanced ECG can better predict LVH and LVSD, it makes sense that it could also better identify individuals at risk of dying compared to a standard ECG interpretation.
An abstract presented at the last American Heart Association meeting in November 2019 reported that the ability of a DNN-analyzed twelve-lead ECG to predict one-year mortality was superior to standard ECG analysis.9 The researchers trained a DNN to predict one-year mortality directly from 1,775,926 ECGs and compared the DNN performance to standard ECG measures — diagnostic patterns and standard measurements.
Three cardiologists separately reviewed the 297,548 ECGs from patients who died that were read as normal. The researchers stated that “The patterns captured by the model were not visually apparent to cardiologists, even after being shown labeled true positives (dead) and true negatives (alive).”
The potential for AI-enhanced analysis of ECG to improve patient care appears promising, but the technology needs further clinical validation from additional studies conducted on varying populations before it is ready to become a useful clinical adjunct to the practicing cardiologist.
1. The computerized ECG: Friend and foe. The American Journal of Medicine. https://www.amjmed.com/article/S0002-9343%2818%2930853-2/fulltext. Last accessed August 2021.
2. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature. https://www.nature.com/articles/s41591-018-0268-3. Last accessed August 2021.
3. A deep neural network learning algorithm outperforms a conventional algorithm for emergency department electrocardiogram interpretation. Journal of Electrocardiology. https://www.sciencedirect.com/science/article/abs/pii/S0022073618302292?via=ihub. Last accessed August 2021.
4. A deep neural network for 12-lead electrocardiogram interpretation outperforms a conventional algorithm, and its physician overread, in the diagnosis of atrial fibrillation. IJC Heart & Vasculature. https://www.sciencedirect.com/science/article/pii/S2352906719301241. Last accessed August 2021.
5. Comparing the performance of artificial intelligence and conventional diagnosis criteria for detecting left ventricular hypertrophy using electrocardiography. EP Eurospace. https://academic.oup.com/europace/article-abstract/22/3/412/5652054?redirectedFrom=fulltext. Last accessed August 2021.
6. Detection of Hypertrophic Cardiomyopathy Using a Convolutional Neural Network-Enabled Electrocardiogram. Journal of the American College of Cardiology. https://www.jacc.org/doi/full/10.1016/j.jacc.2019.12.030?_ga=2.200274285.2130491331.1586866456-722400083.1586006686. Last accessed August 2021.
7. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nature Medicine. https://pubmed.ncbi.nlm.nih.gov/30617318/. Last accessed August 2021.
8. An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction. The Lancet. https://www.thelancet.com/journals/lancet/article/PIIS0140-6736%2819%2931721-0/fulltext. Last accessed August 2021.
9. Abstract 14425: Deep Neural Networks Can Predict 1-Year Mortality Directly From ECG Signal, Even When Clinically Interpreted as Normal. Circulation. https://www.ahajournals.org/doi/10.1161/circ.140.suppl_1.14425. Last accessed August 2021.T