In a recent study published in eClinicalMedicine, researchers developed, trained, and tested three independent artificial intelligence (AI) models to diagnose atherosclerotic cardiovascular disease (ASCVD). Each model was developed to evaluate a separate stage of the ASCVD pathway: elevated coronary artery calcium (CAC), obstructive coronary artery disease (CAD), and regional left ventricular akinesis.

Results highlight that each model correlated well with known clinical ASCVD risk factors and high overall performance. When ensembled together, these AI models could identify at-risk individuals over timeframes as brief as three years, significantly outperforming current diagnostic tools.

ASCVD diagnoses and the potential for ECG-AI

Atherosclerotic cardiovascular disease (ASCVD) is one of the most prevalent non-communicable conditions in the world and the leading cause of mortality globally. It is estimated to be responsible for approximately 50% of all deaths in Westernized countries. ASCVD is thought to be caused by a combination of biochemical blood perturbations, including cholesterol, glucose, homocysteine, and fibrinogen.

The best clinical outcomes of ASCVD involve disease risk detection and the initiation of therapeutic interventions such as lipid-lowering therapy. However, these interventions rely on early risk detection involving expensive non-routine diagnostic tools, which evade a large proportion of the human populace.

Conventional, evidence-based ASCVD diagnosis relies on evaluations of a patient's clinical and demographic information. The most popular are pooled cohort equations (PCE) that use clinical and demographic data to provide a 10-year ASCVD risk assessment.

Unfortunately, PCE and other conventional risk estimators are generalist in their approach, making them error-prone when used on individuals with risk factors outside their predefined parameters. Coronary artery calcium (CAC) scoring has partially addressed this limitation. When used in tandem with conventional risk estimators, CAC reclassification significantly improves estimator accuracy. Since CAC is still a novel tool, it remains expensive, not covered under most healthcare plans, and underutilized despite extensive literature on its robust clinical utility.

The growing success of artificial intelligence (AI) models in clinical and non-clinical scientific research has prompted studies leveraging AI for ASCVD predictions using electrocardiogram data (ECG-AI). Though limited in number, these studies have highlighted AI's good performance and discriminatory capacity to evaluate when a CAC scoring is required versus when it is not.

Research has found that machine learning (ML) models supplied with ECG, CAC, and clinical risk factors outperform currently accepted analyses of each modality assessed independently in estimating cardiovascular risk.

These findings suggest that AI may present untapped value for cardiovascular health assessment without requiring specialized equipment and expensive, invasive clinical procedures.

About the study

In the present study, researchers developed, trained, and validated three independent AI models, each targeting a separate facet of the coronary artery disease (CAD) spectrum. Data for the models was obtained from a deidentified digital electronic health records (EHR) database comprising 7,116,209 patients under treatment from over 70 Mayo Clinic hospitals across 5 US states – Arizona, Iowa, Florida, Minnesota, and Wisconsin.

All patients above the age of 18 years who had records of at least one digital 10-s, 12-lead ECG were included in the study. Models were trained to evaluate CAC scores, obstructive CAD, and regional left ventricular (LV) akinesis, respectively. Each model used 60% of the available data for training, 5% for validation, and 35% for testing. Datasets were split at the patient level to ensure data integrity.

Computation of 10-year ASCVD risk was carried out using the pooled cohort equation (PCE). Analysis variables included high-density lipoprotein levels, systolic blood pressure, total cholesterol, smoking status, age (at the time of ECG), and race/ethnicity. If applicable, clinically diagnosed hypertension was also included. EHR data was additionally queried for mortality, including its underlying cause.

All models were convolutional neural networks (CNNs) with architecture and training method choices deriving from previous research with successful outcomes. The validation dataset (5% of total patient data) was used for hyperparameter tuning. Model input data comprised a 10-s ECG signal with 12 leads. Each lead contains 5,000 samples (500Hz).

"Each model is initialized with pre-trained weights derived from an independently trained self-supervised learning (SSL) model; the objective of this SSL model was to learn a generic contextual representation of ECGs, and it was trained on a dataset consisting of 6 million ECGs".

Model performance testing utilizing the 35% testing dataset was carried out using a bootstrapping approach to compute performance metrics. These metrics included sensitivity, specificity, and the area under the receiver operating characteristics (AUROC). Rate ratios, chi-square tests, Cohen's D, and Cox proportional hazards ratios were used for statistical analyses of risk.

Study findings

All three models present excellent predictive performance when used independently. The CAC score model was found to successfully discriminate between patients with CAC ≥ 300 and those without coronary calcification with an AUROC of 0.88, a sensitive of 78.7% and a specificity of 81.6%. The obstructive CAD model identified patients with CAD with an AUROC of 0.85, a sensitivity of 70%, and a specificity of 81.8%. The regional akinesis model achieved performance metrics of 0.94 (AUROC), 82.2% (sensitivity), and 92.1% (specificity).

The correlations of these models with current laboratory and clinical ASCVD risk factors were similarly high. When the three models were ensembled into a single model, the new model performed better than the sum of its components.

We found that the three models, when ensembled into a single model, combined to provide additive information to a patient's standard-of-care PCE-based cardiovascular risk assessment using a routine and affordable ECG test. Most strikingly, ECG-AI identified patients with elevated acute coronary event risk over timeframes as short as 3 years, even within cohorts already stratified by 10-year ASCVD risk".

These performance gains may be due to interactions between the underlying neural networks of each model, which, when combined, can identify complex CAD-associated patterns via ECG data alone, even for patients without a prior ASCVD diagnosis. When comparing ECG-AI output with the current PCE diagnostic methodology, patients identified as 'at risk' by any of the AI models were also shown to have significantly higher PCE-assessed risk, suggesting that ECG-AI can be used as an adjunct in the clinical setting.

The ability of ECG-AI to predict MI risk in timeframes as short as 3-years is especially compelling because it may prompt more aggressive, collaborative decision making to definitively diagnose and treat high risk ASCVD".

Conclusions

In the present study, researchers developed three independent AI models, each targeting a separate facet of the ASCVD spectrum. The models were trained to use only patient ECG data to predict ASCVD risk. All models showed excellent performance and robust reliability on comparisons with real-world data. When ensembled together, the resultant model was able to predict ASCVD risk in as short a timeframe as three years.

In summary, these data demonstrate that ECG-AI has discriminant value at multiple points along the spectrum of coronary disease. ECG-AI developed to identify 1) elevations in CAC, 2) Obstructive CAD, and 3) Regional LV Akinesis as a marker of possible prior MI are complementary and identify unique risk profiles. Data suggest that ECG-AI designed to detect CAD may be able to support and inform clinical decision making by adding another dimension of easily obtainable, point-of-care data to ASCVD risk assessments, enabling providers to titrate the strength or speed of subsequent intervention accordingly."

Written by

Hugo Francisco de Souza

Hugo Francisco de Souza is a scientific writer based in Bangalore, Karnataka, India. His academic passions lie in biogeography, evolutionary biology, and herpetology. He is currently pursuing his Ph.D. from the Centre for Ecological Sciences, Indian Institute of Science, where he studies the origins, dispersal, and speciation of wetland-associated snakes. Hugo has received, amongst others, the DST-INSPIRE fellowship for his doctoral research and the Gold Medal from Pondicherry University for academic excellence during his Masters. His research has been published in high-impact peer-reviewed journals, including PLOS Neglected Tropical Diseases and Systematic Biology. When not working or writing, Hugo can be found consuming copious amounts of anime and manga, composing and making music with his bass guitar, shredding trails on his MTB, playing video games (he prefers the term ‘gaming’), or tinkering with all things tech.