Artificial Intelligence-enabled Cardiovascular Diagnostic Tools, Techniques, and Methodologies
A new era in Electrocardiogram Analysis
The application of AI to the electrocardiogram (ECG) has seen significant advances recently and has developed in the following two broad categories: (i) tools to automate ECG interpretation, extending human capabilities via massive scalability, important as mobile form factors permit signal acquisition and (ii) algorithms to identify conditions not visible to human readers by training networks to identify multiple, complex, non-linear patterns in the ECG signal to find occult disease (confirmed using other tests such as imaging), or impending disease. In contrast to automation tools in which a human overread provides a gold standard, algorithms identifying occult or future conditions require additional patient information.
Several groups have used large, labelled data sets to train neural networks to accurately apply diagnostic codes to single-lead and multiple-lead ECGs. Hannun et al. used 91 232 single-lead ECGs from a wearable patch to train a network to provide 12 rhythm classes and found that the network outperformed the average cardiologist's read. Subsequently, two mega trials using smart watches based on photoplethysmography technology enrolled 419 297 and 246 541 patients to screen for atrial fibrillation (AF) in under 9 months.[2,3]
These trials confirmed the ability to massively enrol subjects and acquire data, at the cost of high rates of early dropout and a low yield of disease (<0.5% in both studies), and with limited clinical characterization of the study subjects. Ongoing trials will assess these tools in the context of patients selected for arrhythmia risk. Finally, there have been recent reports of interesting research that aimed to develop and validate an AI-enabled ECG algorithm capable of comprehensive 12-lead ECG analysis comparable to that of practising cardiologists.
Furthermore, the AI-ECG has identified occult and manifest cardiac conditions, including ventricular dysfunction, peripartum cardiomyopathy, amyloid heart disease, and pulmonary hypertension, as well as non-cardiac conditions such as hyperkalaemia and cirrhosis.9,10 In addition, special algorithms have been used for the early diagnosis of valvular diseases such as asymptomatic or oligosymptomatic severe aortic stenosis and mitral regurgitation,[11–13] left ventricular hypertrophy,[14,15] myocardial infarction,[16,17] and a number of other conditions. Common findings in these studies include a strong clinical performance [area under the curve (AUC) often >0.90] and detection of disease months to years ahead of the clinical diagnosis.
The significance of these findings remains to be evaluated, taking into account the scalability of electrocardiography, and hence the contribution of AI to its further and more substantial utilization.
The ECG is an ever-present diagnostic tool that has served medical practitioners for more than a century. With the support of deep learning AI techniques, it is clearly entering a new era, in which it may prove to be a powerful detector of subclinical and clinical cardiac diseases, going beyond the boundaries of human observation. There can be no doubt that when the previous capabilities of the ECG are combined with the evolving features of wearable devices such as smartphones, the chances of a much broader and pluralistic diagnostic process will increase rapidly.
Artificial Intelligence-electrocardiogram and Clinical Trials
Clinical trials are essential to demonstrate the ability of novel digital tools like the AI-ECG to improve human health. Factors to consider in evaluating the quality of AI-ECG studies are listed in Table 1. A framework for the assessment of how well AI-ECG clinical trials can predict meaningful outcomes, based on whether the trials are single-centre or multicentre, prospective or retrospective, is shown in Table 2. It is likely that Level 3 or higher would be required for regulatory approval, allowing for variation in specific tests and regional differences. There is a pressing need for additional clinical trials to assess AI-ECG tools. A search of clinicaltrials.gov on 8 October 2021, for trials utilizing the terms 'artificial intelligence' and 'ECG' returned 27 studies, with only 5 completed.
The first AI-ECG prospective trial published, the Eagle study, demonstrated how digital, pragmatic trials can effectively and rapidly enrol subjects, and how the AI-ECG can positively impact clinical practice. It randomized 120 primary care teams from 45 clinics or hospitals in Minnesota and Wisconsin to an intervention arm (clinicians have access to AI-ECG results screening for left ventricular dysfunction when routinely ordering a clinical ECG) or a control arm (no AI results). Despite the development of the pandemic, >22 000 patients were enrolled in 8 months, and the AI-ECG increased the diagnosis in the overall cohort [odds ratio (OR) 1.32, P = 0.007]. The test performance (AUC = 0.92) matched that of the initial retrospective cohort (0.93). Interestingly, the overall utilization of echocardiography was similar in both groups but in the intervention group, more echocardiograms were ordered for patients with a positive AI-ECG (38.1% control vs. 49.6% intervention, P < 0.001), suggesting that the AI-ECG did not lead to more echocardiograms, but to better selection of patients to undergo imaging.
Imaging has been the frontrunner in the application of AI in healthcare, because of the repetitive nature of imaging processing and evaluation. Artificial intelligence may improve imaging quality—and thereby scan and dose time—and assist in segmentation, processing, and analysis. Furthermore, most data are retrieved from a single standardized data source, making it more accessible for large-scale analyses. During the pandemic, critics were pointing out that, despite massive efforts, AI had no impact on the care of COVID-19 patients, while simple straightforward randomized controlled trials did save lives. However, this clearly shows only one side of the coin. The pandemic led to a greater burden on radiology resources, as computerized tomography (CT) scans were carried out routinely in all patients. Artificial intelligence is key in all parts of the imaging pipeline, including acquisition, processing, and analyses.22,23 Furthermore, a plethora of papers have been published during the pandemic, showing the prognostic value of calcium score measurements in COVID-19 chest CT scans.
Those measurements can be automated using deep learning, providing clinicians with information, not only about the pulmonary status of COVID-19 patients, but also their cardiovascular risks. Artificial intelligence will enable automated analyses of routine chest CT examinations for opportunistic cardiovascular screening, allowing early preventive treatment. All these developments, together with the notable Food and Drug Administration clearance of a new technology to identify strokes on brain CT scans enabled by AI, hold out the prospect of a bright future in medical diagnostics.26,27
Retinal Photography to Detect Cardiovascular Disease
Another imaging application that can determine risk across a wide range of diseases is retinal photography. Retinal photography is a non-invasive imaging modality that aides in the diagnosis and treatment of major eye diseases, but can also provide information on the human vasculature and therefore cardiovascular disease. Prior manually coded studies have shown that retinal vascular abnormalities are predictive for cardiovascular disease. Deep learning can extend this knowledge through the automation and detection of more subtle signs that are not clearly visible to the human eye. Several large-scale studies have been published recently, focusing on the predictive value of features extracted from retinal photographs. Studies have shown that deep learning algorithms can predict levels of biomarkers such as haemoglobin to detect anaemia, as well as age, sex, body composition, and creatinine levels, although external validation is warranted before this can be widely adopted in population screening. Another interesting study investigated the predictive capability of deep-learning-enabled coronary artery calcium (CAC) scores derived from retinal scan data. Computerized tomography scans and retinal measurements were performed on the same day and the score derived from retinal images showed an AUC of 0.74 for predicting CAC > 0. Although higher than other single risk factors, such as age, sex, and cholesterol, the added predictive value in the multivariable clinical model was limited (AUC from 0.782 to 0.784). However, the CAC score derived from retinal scans showed a similar performance in predicting cardiovascular outcomes to CAC measured by CT scan (both AUC 0.71). Furthermore, the authors showed in the UK Biobank that this retinal-based CAC score could improve risk stratification in those with borderline or intermediate risk.
However, this method has certain disadvantages. Home-based tests are not yet available, and images with poor quality were excluded in the reported analyses, which is likely to limit the external validity. Real-world data are necessary to estimate the added value in population screening, and the development of mobile applications for self-tests is needed before implementation on a large scale. These deep learning applications are, however, already useful in those who already undergo regular retinal scans, such as diabetic patients, to screen for retinopathy.
To end this section, at least a brief mention should be made of the diagnostic capability and cost-effectiveness of the combined imaging approach, where the use of AI and magnetic resonance imaging yields the atheroma index of the coronary arteries or peripheral vessels as a byproduct of the primary diagnostic evaluation of other organs.
Automation of Imaging Processing
While the application of AI in cardiovascular imaging for clinical decision-making is still in its infancy, the use of AI to automate imaging processing in other fields, such as ophthalmology as discussed above, oncology, and dermatology, has already matured. However, several promising studies using different imaging modalities have recently been published and have shown that cardiology is able to catch up with the other disease domains. A large international collaborative study showed that the coefficient of variation in measuring left ventricular wall thickness by cardiovascular magnetic resonance was significantly lower for machine learning in comparison to human experts. This study involved a cohort of patients with hypertrophic cardiomyopathy, where variations in wall thickness measurements directly impact clinical decision-making by affecting the calculation of sudden death risk and thereby the indication for preventive implementation of an implantable cardioverter defibrillator (ICD).
Another recent example of automation is the International Society for Heart and Lung Transplantation's grading of endomyocardial biopsies in heart transplant patients. The authors compared histological grading performed by expert pathologists with a computer-assisted automated pipeline and showed similar performance of the Computer-Assisted Cardiac Histologic Evaluation (CACHE) grader in comparison to the pathologist (Figure 1). Moreover, they showed only limited attenuation of the performance when it was applied to an external validation data set, indicating good generalizability across different scanning and tissue preparation protocols. International collaborative efforts in the field of transplant research have been hampered by variations in grading by individual centres, which increase the noise-to-signal ratio in the detection of biologically meaningful results when data sets from individual centres are merged. CACHE-enabled automated grading can play an instrumental role in advancing the field of transplant research.
An overview of the 'Computer-Assisted Cardiac Histologic Evaluation (CACHE)-Grader' multicentre validation experiment. The CACHE-Grader's performance was compared with both the grade of record and independent pathologists performing re-grading, demonstrating non-inferiority to expert pathologists, generalizability to external data sets, and excellent sensitivity and negative predictive value. Reproduced by permission from Peyster et al.36
Finally, AI will increasingly be applied in the field of echocardiography. Prior studies have shown that AI can identify different echo views, segment cardiac structures, estimate ejection fraction,37,38 and diagnose diseases such as cardiac amyloidosis. Recently, a study from Stanford also showed that deep-learning algorithms are able to detect pacemaker or ICD leads and, interestingly, are able to predict age, sex, height, and weight based on echo images. Furthermore, they used gradient-based sensitivity mapping methods to highlight the regions of interest for human interpretation. Visualization methods to unlock the so-called black box algorithms are essential if healthcare professionals are to fully adopt the results generated by AI models. These algorithms will support untrained professionals with the interpretation of echocardiograms when cardiological expertise is of limited availability. A recent study showed that deep learning can even help untrained nurses to perform limited echocardiograms for standard evaluation of the left and right ventricular size and pericardial effusion, enabling the use of echocardiograms in non-cardiological settings, such as primary care, COVID wards, or remote areas. However, before its widespread implementation, additional studies regarding safety and generalizability are warranted.
Eur Heart J. 2022;43(4):271-279. © 2022 Oxford University Press
Copyright 2007 European Society of Cardiology. Published by Oxford University Press. All rights reserved.