AI-Enabled ECG in the ED May Pinpoint Dyspnea Patients With Heart Failure

By Marilynn Larkin

August 13, 2020

NEW YORK (Reuters Health) - Applying artificial intelligence (AI) to an electrocardiogram (ECG) may rapidly identify left ventricular systolic dysfunction in patients presenting to the emergency department (ED), an observational study suggests.

"An artificial intelligence tool has the ability to rapidly provide physicians with additional diagnostic information in an acute care setting," Dr. Demilade Adedinsewo of Mayo Clinic in Jacksonville told Reuters Health by email. "It has the potential to improve efficiency in the ED, provides an opportunity to link patients early to appropriate cardiovascular care and also adds incremental valuable information as a screening tool in community EDs or areas without ready access to echocardiography or cardiologists."

"Prospective studies are needed to evaluate the effect of the AI-enabled ECG on long-term clinical outcomes," she added. "Our research team is currently in the process of doing this."

As reported in Circulation: Arrhythmia and Electrophysiology, Dr. Adedinsewo and colleagues retrospectively applied a validated AI-ECG algorithm for the identification of LVSD (defined as left ventricular ejection fraction of 35% or less) in 1,606 patients with dyspnea evaluated in the Mayo Clinic ED. Patients were included if they had at least one standard 12-lead ECG acquired on the date of the ED visit and an ECG performed within 30 days of presentation.

Overall, the median age was 68, about half were female (47%), and the majority were white (91%). The median time to ECG following the ED visit was one day. At the index visit, only 54% had natriuretic peptides (NT-Pro BNP) levels assessed, 43% had high-sensitivity troponin values, and 63% had serum creatinine.

The AI-ECG algorithm identified LVSD with an area under the curve of 0.89 and accuracy of 85.9%. Sensitivity was 74%; specificity, 87%; negative predictive value, 97%; and positive predictive value, 40%.

To identify an ejection fraction < 50%, the AUC, accuracy, sensitivity, and specificity were 0.85, 86%, 63%, and 91%, respectively.

Further, NT-Pro BNP alone at a cut-off of >800 identified LVSD with an AUC of 0.80.

The authors conclude, "The ECG is an inexpensive, ubiquitous, painless test which can be quickly obtained in the ED. It effectively identifies LVSD in selected patients presenting to the ED with dyspnea when analyzed with AI and outperforms NT-Pro BNP."

Dr. Adedinsewo said, "The ECG and NT-pro BNP have the highest level of recommendation by the American College of Cardiology and the American Heart Association when evaluating patients with suspected heart failure. The great thing about this AI tool is that it does not require any additional or new test. It uses the ECG already obtained and provides diagnostic information rapidly - information that the managing physician would otherwise not have without an ECG."

Although the tool is available at all Mayo Clinic sites, she noted, additional research is underway to improve its value and implementation. "As such, I do not know when this technology will be more broadly available or approved for routine clinical use."

Dr. Larisa Tereshchenko of Oregon Health and Science University in Portland, coauthor of a related editorial, commented in an email to Reuters Health, "The study adds to the generalizability and utility of the previously developed AI-ECG algorithm. However, I am concerned that the authors did not report the calibration of the model. To demonstrate an accurate prediction, both discrimination and calibration must be assessed."

"Furthermore," she said, "before implementation in clinical practice, the prediction model must be prospectively validated in large and heterogeneous populations and demonstrate an improvement of clinical outcomes."

"The success of AI in the medical imaging field is undeniable," she affirmed. "However, there is a fundamental difference between a unit of input in AI-models analyzing medical images and ECG. One imaging unit (pixel) has a meaningful interpretation: it reflects tissue characteristics. In contrast,...the ECG sample unit does not have a meaningful interpretation. For the success of AI in the ECG field, a meaningful 'ECG unit' has to be defined."

SOURCE: Circulation: Arrhythmia and Electrophysiology, online August 4, 2020.