In order to bring you the best possible user experience, this site uses Javascript. If you are seeing this message, it is likely that the Javascript option in your browser is disabled. For optimal viewing of this site, please ensure that Javascript is enabled for your browser.
Did you know that your browser is out of date? To get the best experience using our website we recommend that you upgrade to a newer version. Learn more.

Machine learning – helping to bring greater precision to cardiology



Across several areas of cardiology, machine learning is changing our understanding of CVD and helping to transform management. Described below are just two of the innovative new studies presented at ESC Congress 2022, which help to highlight the potential for artificial intelligence (AI) to provide solutions to address everyday problems in practice.

Cryptogenic stroke remains somewhat of an enigma. Although atrial fibrillation (AF) is known to be one of the most common causes of cryptogenic stroke, paroxysmal AF is difficult to detect, even with 24-hour Holter monitoring. Yesterday, Doctor Ki-Hyun Jeon (Seoul National University Bundang Hospital - Seongnam, Republic of Korea) and Doctor Joon-myoung Kwon (Medical AI Co. - Seoul, Republic of Korea) presented a novel approach to diagnosing AF using a deep-learning algorithm (DLA) that may help to guide prescribing of anticoagulation therapy.

A DLA was developed to detect AF during sinus rhythm using 12-­lead ECG data obtained from 10,605 adult patients who had at least one AF rhythm in the study period (2016–2021) and 50,522 non-­AF patients who had no reference to AF in the ECG or electronic medical records. The DLA was based on a convolutional neural network, a class of artificial neural network most commonly applied to analyse visual imagery. Following development, the area under the curve (AUC) of the final DLA was 0.862 (95% CI 0.850 to 0.873).

The DLA was then externally validated using a cohort of 221 patients with cryptogenic stroke and implantable cardiac monitoring (mean duration, 15.1 months) among whom AF of more than 5 minutes had been detected in 32 patients (14.5%). In this cohort, the diagnostic accuracy of the DLA to detect AF during sinus rhythm was 0.811 and the AUC was 0.827. The sensitivity, specificity, positive predictive value and negative predictive value of the model were 0.828, 0.808, 0.415 and 0.966, respectively. The authors concluded that their DLA outperformed other conventional predictive methods based on clinical factors, such as CHARGE­-AF, C2hest and HATCH.

Today, results from a machine-learning analysis to predict outcomes in patients with chest pain and normal high-sensitivity troponin (hs-Tn) T or I levels are presented by Doctor Anna Myredal (Sahlgrenska University Hospital - Gothenburg, Sweden). The researchers used data from 9,314 patients admitted due to chest pain, with normal hs-­TnI or hs-­TnT, who underwent angiography but did not receive a final diagnosis of acute myocardial infarction from the SWEDEHEART registry. By studying angiographic findings on a coronary artery segmental level, the researchers developed machine-learning models to predict 1-year survival, risk of future acute coronary syndromes (ACS) and to rule out unnecessary angiographies (angiography that did not lead to any intervention). Models predicting 1-­year survival and future ACS included 130 candidate predictors, while models for unnecessary angiography included 110 predictors. Approximately 50,000 models were built using gradient boosting, extreme gradient boosting, random forest, artificial neural networks and logistic regression.

In the cohort, 1-­year mortality rate was 0.9%, the 1­-year rate of future ACS was 2.7% and the rate of unnecessary angiography was 61.5%, with up to 40% having normal angiography. A strong association was noted between hs-Tn levels (within the normal range) and severity of coronary atherosclerosis across the segments. For example, 32.4% in patients with hs-­TnI 26–35 ng/L had >50% stenosis in segment 6 compared with 12.6% in those with hs­-TnI 0–5 ng/L. When adjusted for age and sex, the hazard ratio for 1-­year mortality for hs­-TnI 26–35 ng/L versus hs­-TnI 0–5 ng/L was 5.73 (95% CI 2.14 to 15.35). The strongest predictors of 1­-year mortality were C-­reactive protein, body mass index, estimated glomerular filtration rate, age, time from symptom onset to critical care unit admission, systolic blood pressure and hs­-TnI.

The machine-learning models were found to reliably predict 1-­year risk of death, future ACS and unnecessary angiographies – extreme gradient boosting was the best performing model, with AUCs of 0.77, 0.77 and 0.78, respectively, and excellent calibration. The authors also concluded that hs-Tn levels within normal range constitute a strong predictor of all these outcomes and questioned the definition of ‘normal’ hs-Tn.

Learn more about the potential for AI to predict and improve outcomes in cardiology at today’s oral Abstract Session.

The content of this article reflects the personal opinion of the author/s and is not necessarily the official position of the European Society of Cardiology.