Deep learning to improve ECG interpretation

Traditional computerised ECG interpretation uses a set of criteria developed by humans to make classification decisions based on the ECG input. These criteria are generally created from reference ranges in a population. For example, the computer interpretation on an ECG machine can be programmed to measure PR intervals, and where the PR interval is greater than 200 ms, the computer will make the diagnosis of first degree heart block. By contrast, the deep learning approach is able to identify important features from thousands of input ECGs and develop its own set of complex criteria for each diagnosis of interest. This more complex approach can be thought of as being analogous to how a clinician might develop skills in ECG reading over a career of ECG interpretation and clinical correlation. A deep learning approach to ECG interpretation could be more accurate than a traditional computerised interpretation,3 because ECG interpretations by deep neural networks might outperform skilled humans.4

Screening for prevalent disease

Given the relative ease and low cost of performing an ECG, an important potential evolution in the application of AI enabled ECG is detection of prevalent disease. This distinction to the application of interpretation above is because AI enabled ECGs may be able to detect pathology that standard ECG interpretation by a clinician may miss. A prominent example is the impressive performance of AI enabled ECGs for the identification of asymptomatic left ventricular dysfunction (accuracy 85.7%, area under the receiver operating characteristics curve 0.93).5 AI enabled ECGs have also been applied for aortic stenosis detection6 (0.85). The ubiquitous ECG also provides an opportunity for screening for rarer diseases that may otherwise go undetected in current clinical practice. Hypertrophic cardiomyopathy by AI enabled ECG detection has achieved an impressive performance7(sensitivity 87%, specificity 90%, 0.96). AI enabled ECGs have also been applied for amyloidosis (0.91)8 and peripartum cardiomyopathy detection9 (area under the receiver operating characteristics curve of 0.87-0.92 depending on ejection fraction cut-off). These deep neural networks are able to identify both features that are apparent to humans and novel features that had not previously been described.10 Application of AI enabled ECGs for screening could be particularly relevant in primary care, where diseases currently may go undetected and could trigger referral on for further testing. However, we note several potential limitations. Any screening approach would need to be applied to appropriate populations, for example, opportunistic screening for asymptomatic left ventricular dysfunction in the general population may be appropriate given the high prevalence of the condition and the highly effective treatments that are available. By contrast, screening for rarer diseases without well established treatments, such as amyloidosis, is less likely to be beneficial. The potential harms of screening (including overdiagnosis and false positive diagnoses generating unnecessary additional testing) must be carefully considered and this approach needs evaluation in prospective clinical trials. Additionally, AI enabled ECG interpretation must be evaluated in an appropriate clinical context. For example, AI enabled ECGs may report evidence of aortic stenosis, but this finding may be incidental and not the cause of the patient’s symptoms.

Prediction of future disease

The ECG is generally used by clinicians to identify prevalent abnormalities. AI enabled ECGs, however, have the potential to identify patients at risk of future disease. A potentially impactful application is the prediction of future atrial fibrillation from the sinus rhythm ECG (30 day prediction,11 area under the receiver operating characteristics curve 0.87; five year prediction, 0.909).12 This prediction could guide targeted screening, such as prolonged rhythm monitoring or mobile health devices. Another application of AI enabled ECG is to predict sudden cardiac arrest within 24 h13 (area under the receiver operating characteristics curve of 0.948 in an external dataset using 12 leads and 0.925 using a single lead I ECG). This screening might have the potential for use in a wearable device or implantable loop recorder in patients who are at an increased risk of cardiac arrest, but not sufficiently high risk enough to warrant implantable cardioverter defibrillator implantation. Although promising, further work is needed to evaluate potential applications of AI enabled ECG in this context.

Disease phenotyping

Use of the ECG for unsupervised learning has potential to identify important disease phenotypes that may not be apparent from human inspection of clinical data and the ECG. Currently patients are selected for cardiac resynchronisation therapy comparatively crudely, by identification of left or right bundle branch block and measurement of QRS duration. Unsupervised machine learning has been applied to the raw QRS waveforms of patients who had cardiac resynchronisation therapy devices implanted for heart failure. Two prognostically important subgroups were identified, with importance beyond the crude parameters that we use today. The findings indicate unsupervised machine learning may be used to better identify patients who would benefit from cardiac resynchronisation treatment.14 Unsupervised machine learning, in the form of a convolutional autoencoder, has also been used to predict nocturnal hypoglycaemia.15 An autoencoder (depicted in figure 1) attempts to compress the input signal into a latent representation of features (in this case, 20 or 50 neurons) and then decompress to create a signal as close to the original as possible.2 Through this technique, the latent representation provides features that describe the input signal and can then be used for clustering or for a supervised machine learning task. The algorithm correctly identified clusters of ECGs based on glucose concentrations with 90% accuracy.

Prospective validation of AI enabled ECG in clinical practice

Applications of AI enabled ECG have been extensively studied in retrospective datasets; however, prospective validation in clinical environments is important before clinical adoption. The first randomised clinical trial of AI enabled ECGs versus usual care was completed in 2021.16 They applied the AI enabled ECG algorithm for asymptomatic left ventricular dysfunction detection previously described5 in a cluster randomised trial of 120 primary care teams (22 641 patients). Overall, the AI enabled ECG group had a 32% increased rate of low ejection fraction diagnosis (odds ratio 1.32 (95% confidence interval 1.01 to 1.61), absolute diagnosis 2.1% v 1.6%) compared with usual care. AI enabled ECG could potentially be used in primary care to guide clinicians as to which patients should be referred for echocardiography. Importantly, echocardiogram use was not significantly different between the groups, indicating that AI enabled ECG may be a cost-effective tool, although formal cost effectiveness analysis is required.

An observational study published in 2022 reported on the application of a deep neural network for left ventricular dysfunction screening embedded within an ECG enabled stethoscope.17 By use of an ECG enabled stethoscope, participants had a single lead ECG recorded, the area under the receiver operating characteristics curve was 0.85 for detecting left ventricular ejection fraction of less than 40%. Embedding AI enabled ECG within the familiar stethoscope may further increase its usability through quick recordings that can be taken in all patients who are seen in person. Figure 3 shows how AI enabled ECG may be used in the near future to support decision making for clinicians by providing suggestions for further investigation and management in both the outpatient and emergency settings. Despite the great potential of AI enabled ECG, improved clinical outcomes have not been shown; future studies will need to address this.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

A depiction of how AI enabled ECG may be used in the near future to guide tailored further investigation and management in both outpatient and emergency settings. Examples of potential applications of AI enabled ECG prediction in each setting are given, with how these predications may guide clinical management

Protecting human autonomy

The increasing role of AI in many fields, not just healthcare, risks the transfer of decision making power to machines. In healthcare, the risks are far too high, and many believe that humans must always have the final decision. The examples discussed in this review are predominantly about screening or identifying disease. In these scenarios, AI enabled ECG acts to enhance disease detection, where conditions could otherwise be missed, and in each case, a confirmatory test would generally be required before making any major changes to treatments. The potential exception to that is with respect to atrial fibrillation prediction. AI enabled ECG could be used to recommend pre-emptive anti-coagulation where the predicted risk of incident atrial fibrillation is particularly high. Any such workflows would need to be supported by clinical trials showing such a practice is beneficial and should always involve human input and shared decision making.

Ensuring transparency, explainability, and intelligibility

One area of concern with deep learning models is the challenge in understanding reasons for the decision recommended by the model. Several methods, including saliency mapping, can partially overcome these limitations. A saliency map is an image where the brightness or colour of each pixel indicates how much of an impact that area of the image or signal has on predicting the outcome class. For example, a saliency map for a neural network designed to identify left bundle branch block would be expected to highlight the QRS of an ECG. However, no perfect method is available for explainability and all methods have important limitations.19 For example, saliency mapping is best suited to identifying incorrect foci for models rather than truly providing explainability. Using saliency mapping, a review of studies using AI models to detect covid-19 from a chest radiograph showed that many models used erroneous features.20 For example, parts of the image outside the lung fields such as the portable marker, which would generally reflect a more unwell patient, were used by the models to predict adverse outcomes.

Another thought is that interpretability is not required for adoption of AI models into clinical practice. In striving for explainability in our AI models, some people argue that we may limit machines to the level of reasoning that humans are capable of, and therefore deprive the world of the unique problem solving capabilities of AI.21 Furthermore, AI is held to higher standards than other tools used by clinicians. Several examples exist of medical treatments and procedures where the exact mechanism of benefit is not clearly understood but treatments are used extensively after favourable data was available from clinical trials.22 23 Human clinical decision might often escape explainability,24 so why should AI be held to this potentially unrealistic target? Human clinical decision making may also be influenced by unmeasured factors relating to human interaction, such as facial expressions that may convey anxiety or degree of pain. These factors would not be considered by most current AI models. The answer to this question is likely a compromise between these two extremes—some explainability in AI is necessary; however, requiring complete understanding of the AI model is probably an unrealistic goal that will stifle innovation if pursued rigorously.

Ensuring inclusiveness and equity

AI has the potential to both widen and close existing forms of bias and discrimination. Examples from image analysis from large datasets have shown poor performance on images of people from ethnic minorities, due to the largely white population used for training the AI model.25 On the contrary, other studies have shown that AI enabled ECG may in fact perform equally well for people from ethnic minorities, even when they did not comprise a large proportion of the training cohort.17 26. Rigorous testing must be used to ensure systemic biases are not introduced through application of AI enabled ECG.

The present situation

AI enabled ECG is not currently ready for clinical use, however, prospective clinical studies have shown much promise. Figure 3 depicts potential applications for AI enabled ECGs in both the outpatient and emergency settings. The earliest clinical applications of AI enabled ECGs are likely to be additive to current clinical pathways, for example, in screening for asymptomatic left ventricular dysfunction or for patients at high risk of occult atrial fibrillation. Yet, substantial work remains in order to fully realise the potential of AI enabled ECG.