Artificial Intelligence-Enabled ECG: Physiologic and Pathophysiologic Insights and Implications

INTRODUCTION

Over one century removed from its birth, the electrocardiogram (ECG) remains an essential, non-invasive, and cost-effective diagnostic tool to date. ECG interpretation competency requires a solid knowledge base of electrophysiologic principles coupled with experience. Accurate ECG interpretation can provide insights into an individual’s electrocardiac activity, including the underlying rhythm, structural changes, and even evidence of evolving or prior myocardial injury. Its ability to instantaneously evaluate cardiac electrical activity has shaped and continues to inform medical management.

In recent years, technological advances have given this diagnostic tool a new dimension. Advances in computational power alongside the simultaneous exponential growth in available digitized data have led to the development of deep learning models in various fields spurring tremendous promise of clinical application. Since ECGs are recorded with standard protocols and can be archived into usable digital formats, the electrocardiographic signal displayed naturally serves as an excellent substrate for deep learning models.

Deep learning models can be trained to be able to recognize patterns consistent with various rhythms. Recent work has shown their capabilities to extend beyond simple rhythm analysis, but to also include the detection of reduced ejection fraction (1), assessment of serum potassium levels (2), prediction of age and sex (3), prediction of atrial fibrillation (4), and identification of unique cardiac pathologies such as hypertrophic cardiomyopathy (5). Perhaps even more exciting, these models are capable of improving their recognition of various unique patterns making their potential diagnostic yield unparalleled.

Despite the exciting developments of numerous artificial intelligence-enabled ECG (AI-ECG) models capable of detecting specific and unique electrocardiographic patterns, numerous questions remain unanswered. Perhaps one of the most glaring of these questions to many is: what are these models detecting that the human eye is missing? In this review, we highlight some of the recent developments in the field and their potential clinical implications, while also attempting to shed light on the physiologic and pathophysiologic features that enable these models to have such high diagnostic yield.

Atrial fibrillation

Atrial fibrillation (AF) is estimated to affect over 30 million people around the world (6,7). These patients are at increased risk of stroke, heart failure, and mortality (8,9). Thus, identification of these individuals is important as medical intervention (e.g., anticoagulation therapy, left atrial appendage closure, and/or ablation procedures) can reduce the risk of thromboembolic events, hospitalizations, and mortality (10,11).

Unfortunately, AF remains underdiagnosed and screening remains a challenge. The fleeting nature of paroxysmal AF makes it difficult to capture on a single standard 10-second 12-lead ECG. While prolonged cardiac monitoring may detect AF in more individuals, this approach is highly resource intensive and expensive. Furthermore, clinical risk scores to identify at-risk individuals are not perfect.

A recent AI-ECG algorithm has shown tremendous promise in the ability to detect patients with a high likelihood of paroxysmal AF or atrial flutter on a standard 12-lead ECG demonstrating sinus rhythm (area under the receiver operating curve [AUC] 0.87) (4). Its diagnostic prediction even improved in those with multiple ECGs (AUC 0.90). This demonstrates a potential means of identifying those individuals missed during routine testing, who are at risk and may benefit from therapy.

While it is presumed that the convolutional neural network’s (CNN’s) exhaustive analysis of the ECG is detecting subtle perturbations in the electrical signal that would otherwise be missed by the human eye, it is still unclear what these subtle features are. Many studies have found normal sinus rhythm to not be a reliable predictor of atrial function, so it does not seem unreasonable to assume that some electrocardiographic precursor signature of AF exists. In fact, one study reported that nearly one-third of patients with AF undergoing cardioversion had non-sinus contraction of the left atrial appendage despite their surface ECG demonstrating normal sinus rhythm (12). Others have shown interatrial block to correlate with the risk of incident AF and stroke (13,14). Additionally, P wave dispersion has been demonstrated in multiple studies to predict new onset AF (15-17). Thus, as aforementioned, the presence of a non-sinus electrocardiographic precursor signature of AF despite the presence of sinus rhythm on the surface ECG is not implausible. Whether there are non-sinus electrocardiographic features from structural changes preceding AF (e.g., fibrosis, myocyte hypertrophy, chamber enlargement) or other pathophysiologic events remains uncertain.

While this electrocardiographic precursor signature of AF may not be clearly evident to the human eye, the ability of an AI-ECG algorithm to risk stratify patients can potentially result in a paradigm shift in AF management. For instance, in cases of an embolic stroke of undetermined source (ESUS), could the AI-ECG algorithm serve as a surrogate marker and tip the scale to favor those that may benefit from anticoagulation? This is important since empiric anticoagulation without documented AF has demonstrated no additional benefit and potential harm in these patients (18,19). Could the algorithm’s prediction also be used as an indicator of those who may benefit from anticoagulation for primary prevention? It is not difficult to imagine the vast number of clinical applications that a rapid, non-invasive, and cost-effective tool could have in altering AF management and helping achieve arguably the most important objective in AF management – to prevent thromboembolism and its related complications.

Electrolytes

A growing area of interest in clinical care is precision medicine, which is an individualized approach to patient care that shifts away from population-based practice and towards individual patient characteristics (e.g., genetic variability, environment, and behaviors) (20,21). Specific patient characteristics related to clinical outcomes, such as serum electrolytes, allow for deeper insight with regard to organ function.

AI-ECG models have demonstrated promise in detecting serum potassium levels. Models utilizing signal processed single-lead ECGs (22) and ECGs obtained using smartphone technology (24) were able to calculate serum potassium levels non-invasively in a small sample of patients undergoing dialysis treatments. A deep learning model trained on over 1.5 million ECGs from the Mayo Clinic effectively detected hyperkalemia using only two leads from a standard 12-lead ECG (AUC 0.85-0.88) (2).

The ability of an AI-ECG to accurately predict serum potassium levels is thought to be related to subtle ECG features resulting from physiologic changes to myocardial tissue in response to varying potassium levels. Very low or high levels of serum potassium lead to changes in cardiac myocyte action potential, which affects cardiac impulse generation and propagation. These electro-physiologic changes become apparent on the ECG when serum potassium levels exceed a certain threshold. In addition to potassium, other electrolytes (e.g., calcium, magnesium, and sodium) and medications can also alter cardiomyocyte resting membrane potential. These membrane potential variations can alter all cardiac cycle intervals (i.e., PR, RR, QRS, QT and JT intervals), cause ST segment deviation, and the risk of progression to fatal arrhythmias (24). Given our current knowledge of how electrolyte levels can result in characteristic ECG features, it is likely that the AI-ECG models are detecting subtle changes in cardiac conduction before they become overtly evident and detectable by the human eye.

Other machine learning approaches have also demonstrated their effectiveness in identifying acute kidney injury by predicting baseline serum creatinine levels using clinical parameters available within two hours of admission to the intensive care unit (25). This suggests that currently available tools and tests can be enhanced with AI to detect physiologic changes related to electrolyte abnormalities. Moreover, the integration of various AI models portend even greater clinical utility than simply electrolyte level detection, but the potential to enhance disease detection, prognostication, and even prevention.

Age

Physiological age, as opposed to chronological age, is another important measure in precision medicine associated with clinical outcomes. A deep learning model incorporating patient vital signs and laboratory tests has been used to predict discrepancies between with physiological and chronological age, demonstrating larger discrepancies to be associated with poorer health (26). AI-ECG algorithms have also demonstrated the ability to estimate physiological age. One AI-ECG model found that when the predicted age was seven years greater than the chronological age, there was a higher incidence of cardiovascular comorbidities (3). These findings suggested that the AI-ECG age prediction was a measure of physiological health.

In a patient who underwent cardiac transplantation, the AI-ECG algorithm predicted that the age was higher than the chronological age prior to transplant. After cardiac transplant and improvement in co-morbid conditions, the AI-ECG algorithm predicted a lower age that was closer to donor age. This case exemplifies the ability of the AI-ECG to identify age-related physiologic changes on a standard 12-lead ECG. While the actual changes detected on the ECG by the AI-ECG algorithm are unknown, it is likely that other markers of cardiac pathology are recognized. For instance, the presence of left ventricular hypertrophy (LVH) on the ECG is a marker of chronically elevated blood pressures, which frequently correlates with age.

Hypertrophic cardiomyopathy

Hypertrophic cardiomyopathy (HCM) has an estimated prevalence of 1 in 200-500 individuals in the general population (27,28). Echocardiography and MRI are the gold standard methods for the detection of otherwise unexplained LVH (29-31). However, these modalities are not practical or cost-effective for widespread screening of asymptomatic individuals. This is particularly relevant among athletes, in whom a diagnosis of HCM is associated with an increased risk of sudden death. On the other hand, the ECG is ubiquitous and inexpensive. In HCM, the ECG reflects several structural and electrophysiological abnormalities, including LVH, atrial enlargement, myocardial fibrosis, myocardial strain, as well as ventricular depolarization and repolarization abnormalities. An abnormal ECG may be an early marker of HCM in children even before hypertrophy is evident on echocardiography (32), and ECG abnormalities evolve along with the progressive natural history of the disease – ranging from increased QRS voltage and repolarization abnormality typically in childhood and adolescence to atrial enlargement and myocardial fibrosis later in life (33).

More than 90% of patients with HCM have electrocardiographic abnormalities (34), but these abnormalities are nonspecific and can be indistinguishable from LVH. These features include LVH criteria, left axis deviation, prominent Q waves, and T-wave inversions. Some of these changes can be observed in subjects with athletic heart adaptation (35), but also in pathologic conditions other than HCM, such as Fabry’s and Danon’s disease (36). In younger individuals, juvenile T wave inversions may also be mistaken for HCM-related ECG changes. Sets of ECG criteria have been proposed to distinguish between HCM and athletic heart adaptation, but their diagnostic performance has been inconsistent when external validations have been trialed (37, 38). These challenges, along with the relatively low prevalence of HCM in unselected populations, have hampered the enthusiasm for ECG screening for HCM in general populations or in athletic pre-participation settings. However, employing a powerful, fully automated screening approach for AI-ECG-based HCM detection may overcome some of these limitations while concurrently providing valuable insights into HCM physiology.

Using a cohort of approximately 2,500 patients with a validated diagnosis of HCM and over 50,000 non-HCM age- and sex-matched controls, an AI-ECG CNN was recently developed to diagnose HCM based on the ECG alone (5). The CNN model was then tested in an independent testing cohort of 612 HCM and 12,788 control patients, demonstrating an AUC of 0.96 (95% CI, 0.95-0.96) with a sensitivity 87% and specificity of 90%. Model performance was similar in subgroups of patients meeting ECG criteria of LVH, among those with seemingly normal ECGs as well as those with sarcomeric genetic mutations. The performance of the AI-ECG model was superior in younger patients (<40 years), while its performance declined with increasing age.

Figure 1 demonstrates a case of a young woman with HCM who underwent surgical septal myectomy. Her pre-myectomy ECG had relatively minor abnormalities. However, the AI-ECG algorithm indicated an HCM probability of 72.6%. In contrast, her post-myectomy ECG demonstrated an HCM probability of only 2.5% despite more striking ECG findings. This example illustrates that AI-ECG may provide a comprehensive assessment of HCM-related pathophysiology reflected on the ECG that does not rely on a single feature. It can be postulated that the AI-ECG heavily relies on wall thickness – the hallmark of HCM – to determine whether an ECG belongs to someone with HCM or not. However, the surrogate of wall thickness on the ECG is LVH and it is noteworthy that the AI-ECG algorithm in the aforementioned study (5) performed well in distinguishing electrocardiographic LVH due to HCM from LVH unrelated to HCM. Furthermore, it is important to note that neither pre- or post-myectomy ECG in the aforementioned example demonstrate left ventricular hypertrophy or any other distinctive features suggestive of HCM with massive hypertrophy.

Figure 1: AI-ECG model performance before and after septal myectomy.

Artificial intelligence model performance in a 21-year-old woman with obstructive hypertrophic cardiomyopathy (HCM) before (A) and after (B) septal myectomy. Prior to myectomy, the patient had massive septal hypertrophy (30 mm). Adapted with permission from Ko et al. JACC VOL. 75, NO. 7, 2020: 722–33.

Another hypothesis is that the AI-ECG assessment is driven by the LVOT gradient which might be reflected on the ECG through changes related to myocardial strain, depolarization delay, or even left atrial enlargement and fibrosis. This hypothesis was not tested in the previous work (5). However, in ongoing unpublished work we are examining the longitudinal changes in AI-ECG scores in HCM patients treated with a targeted pharmacologic agent. On serial on-treatment ECGs, there was a significant downtrend of the AI-ECG HCM prediction score suggesting reversal of changes related to HCM. The strongest relationship for the decline in HCM score and clinical response was its correlation with the LVOT gradient decrease, suggesting that the LVOT gradient and the multitude of its effects on hemodynamic and electrophysiologic remodeling may indeed be a dominant AI-ECG marker. These findings suggest that a simple and ubiquitous test, like the ECG, enhanced with powerful AI analytics, may provide a means for continuous assessment of complex physiology and treatment monitoring in HCM.

The ability of AI-ECG to detect complex pathophysiology was also demonstrated in a different study where investigators used a large ECG dataset to develop machine learning models for the detection of HCM and for quantitative assessment of cardiac structural details, such as left ventricular mass, left atrial volume, and tissue Doppler velocity of mitral annular early-diastolic excursion. Other conditions such as pulmonary arterial hypertension, cardiac amyloidosis, and mitral valve prolapse were also investigated (39). This study utilized a novel combination of CNNs and hidden Markov models. The performance of the HCM model was quite favorable with an AUC of 0.91. The same investigators also reported good machine learning model performance for the diagnosis of pulmonary arterial hypertension (AUC 0.94), cardiac amyloidosis (AUC 0.86), and mitral valve prolapse (AUC 0.77).

These recent works demonstrate that the discrete ECG findings known to correlate with HCM do not fully explain the CNN performance and therefore suggest that there are further physiologic signals yet to be characterized. Nevertheless, they do reveal how unlabeled data sets can help us uncover new physiologic principles and may allow “deep phenotyping” for both qualitative and potentially quantitative cardiac diagnostics. External and prospective validation efforts of the above tools are underway. Furthermore, the refinement of AI-ECG models to address specific unmet needs of major significance, such as distinguishing HCM from athlete’s heart and predicting sudden cardiac death risk in HCM, will be important in their widespread adoption in clinical practice in screening, diagnostic, prognostic and treatment monitoring settings.

FUTURE WORK AND DISCOVERY

Recent advancements in computational power and availability of digitized data have paved the way for the application of machine learning techniques to the standard 12-lead ECG. AI-ECG algorithms have demonstrated their potential beyond rhythm analysis to include the ability to predict the likelihood of various clinical pathologies and markers. Table 1 highlights recent important works in machine learning related to AF, electrolytes, age, and HCM. While the specific electrocardiographic features the models are detecting remain unclear in some cases, plausible explanations exist for their predictions. As additional AI-ECG algorithms are developed, it will be important for researchers to invest time scrutinizing the accuracies and inaccuracies of their model’s predictions in relation to the apparent electrocardiographic features in order to better understand why such predictions are made. Doing so will help align the model predictions with the underlying pathophysiology.

Table 1:

Recent important works in machine learning related to AF, electrolytes, age, sex, and HCM.

Recent Advances in AI-ECG	PMID Reference
Atrial fibrillation (AF)
Identification of AF from sinus rhythm ECGs (2019)4	31378392
AI-ECG AF model and CHARGE-AF score independently predict incident AF (2020)	33185118
Combination of photoplethysmographic, single-channel ECG, and AF AI-ECG model facilitates AF detection (2020)	32354449
Supervised fully connected artificial neural network can identify AF with portable ECG devices (2020)	32599796
Explainable deep learning model detects AF (2021)	33271204
Explainable AI model for AF using Holter ECG waveforms (2021)	34053998
Electrolytes
Deep learning model detects potassium, sodium, and calcium electrolyte imbalances (2021)	33719135
Deep learning model detects hyperkalemia in patients with renal disease using 2 ECG leads (2019)	30942845
Deep learning model recognizes severe dyskalemias via 12-lead ECG (2020)	32134388
Deep learning model detects hypokalemia via ECG in emergency patients (2021)	34483253
Age and sex
AI-ECG model estimates patient age and sex (2019)	31450977
Deep neural network-estimates ECG age as a prognostic marker for mortality (2021)	34433816
Hypertrophic cardiomyopathy (HCM)
AI-ECG model detects pediatric HCM (2021)	34419527
Machine learning model identifies HCM patients with ventricular arrhythmias (2019)	30952382
AI-ECG model detects HCM via 12-lead ECG (2020)	32081280

In Figure 2, we propose a practical approach to help better understand and discover physiologic patterns detected by AI-ECG models. This proposed approach takes advantage of using known physiologic parameters (e.g., heart rate (HR), PR interval, QRS complex, T wave, etc.) and unknown physiologic patterns detected by the AI-ECG models to improve the diagnostic yield of the models. In addition, we propose that using unknown physiologic patterns may assist in new physiologic discovery via physiology experiments and in silico modelling.

Figure 2: Proposed approach to help better understand and discover physiologic patterns detected by AI-ECG models.

First, physiologic experimentation can help us understand what perturbations affect the model and if there are specific associations (i.e., ECG features/patterns) an AI-ECG model is able to identify. In parallel, in silico (computer) modelling can also help in comprehending and improving AI-ECG models. For instance, saliency maps can provide an understanding as to why a model gives the output it does. They can highlight specific features of an image (in this case, the ECG) that the model focused on to provide its prediction. Additionally, generalized adversarial networks (GANs) can provide a means to improve AI-ECG models. GANs essentially consist of a pair of neural networks – a generator and a discriminator. The discriminator acts to identify true images (in this case, true ECG patterns that fit with the disease state of interest), whereas the generator creates fake images (in this case, fake ECG patterns) in attempt to fool the discriminator into classifying the fake images as real images (i.e., real ECG patterns). The discriminator provides an output (i.e., probability), which can then be used in a back-propagation process to update the weights of both neural networks in an attempt to improve the identifying of an image (i.e., ECG pattern) for a specific disease state. Multiple iterations (i.e., epochs) of real and fake images are used to continuously update the weights to further refine the model’s overall discrimination power. Lastly, the use of adversarial attacks can also help understand a model’s output. These attacks serve as a small, intentional feature perturbations that cause the model to make a false prediction. In doing so, they can help recognize which specific features are contributing the most to a model’s predictions. Thus, in tandem physiologic experience and in silico modelling can help us to use previously unknown patterns detected by the AI-ECG model to determine new physiological processes and their corresponding ECG features.

Unlike conventional computerized ECG interpretation algorithms, AI-ECG interpretation algorithms have the potential to continue to improve their predictions. For instance, reinforcement learning is a trial-and-error training process that can help a model not only achieve a desired goal, but also augment its ability to do so. This ongoing learning capability is an important feature to optimize AI-ECG models for clinical use. Once models are optimized for solving specific task, they then have the potential to be applied to a new application (i.e., transfer learning). A better understanding of the physiology behind the AI-ECG models’ predictions would allow for which next task is best to assign a specific model to.

Aside from physiologic understanding, ethical concerns remain surrounding the application of AI in medicine, particularly with the potential for AI algorithms to replicate or intensify human biases in decision making (51), as well as concerns related to the misclassification of acute pathology that a human would be much less likely to, as seen with adversarial examples (52).

Finally, with the evolution of a plethora of ambulatory cardiac monitoring devices, there is now a means for the ECG to serve as diagnostic and monitoring tool outside of the standard clinical setting. This can have significant advantages to patient care, but also raises new hurdles. For instance, while the cardiac biosignals can be attained outside of the clinical setting, this data can easily be distorted by a number of factors (e.g., environmental noise, electrical interference, concomitant unrelated electrical activity from the body, etc.) making it difficult to use clinically. Additionally, even with clean electrical recordings, most wearable devices do not have comprehensive analytical capabilities, thereby limiting its utility for the typical end user and requiring sole review by a skilled and trained clinician before the recording can be actionable. Given this foreseeable increasing workload to clinicians, systems that can ingest high-quality ECG signals from a variety of devices and utilize AI-ECG models to provide accurate diagnostic results and streamline clinical workflow have been proposed (53,54). While much of this work is in the infancy stages and relies on the diagnostic accuracy of the AI-ECG models, the potential clinical value it can provide patients undeniable. Furthermore, the insights attained by continuous ambulatory cardiac monitoring in conjunction with other biosensors and AI learning capabilities may enable discovery of new physiologic processes.

CONCLUSION

Despite the many foreseeable and unforeseeable hurdles that lie ahead, the potential clinical utility of AI-ECG in patient care is undisputed. The application of AI to a rapid, non-invasive, and inexpensive diagnostic test like the standard 12-lead ECG has demonstrated its ability to transform an old diagnostic tool into a paradigm-shifting tool in patient management. Furthermore, the universal nature of the ECG and technologic advances in wearable and personal monitoring devices may allow for the use of AI-ECG outside of the hospital setting and in resource-limited regions around the world. Finally, merging AI-ECG with other clinical measures and diagnostic tools (e.g., imaging, vital signs) may not only help improve predictions and bring about new discoveries, but also allow for identifying and correlating specific electrocardiographic findings with various disease processes.