ECG AI algorithm identifies left ventricular systolic dysfunction
Patients presenting to the emergency room (ED) with dyspnea have left ventricular (LV) systolic dysfunction, using electrocardiograms analyzed by AI.
Demilade Adedinsewo, MD, principal investigator in the Department of Cardiovascular Medicine at the Mayo Clinic in Jacksonville, Florida, told Healio: "The AI ECG can detect left ventricular systolic function faster and more accurately in patients with shortness of breath than NT-proBNP. Improve and expedite emergency department diagnosis and provide a unique opportunity to identify high-risk cardiac patients earlier and link patients to appropriate cardiovascular care."
patients with breathing difficulties
In the retrospective study, published in Circulation: Arrhythmias and Electrophysiology, researchers analyzed data from 1,606 patients (median age 68; 47% women) between May 2018 and 2019 Difficulty breathing during 2 months. These patients had at least one ECG within 24 hours and within 30 days of their ED visit. Those with previously diagnosed systolic, diastolic, or unexplained heart failure were excluded.
The primary outcome of this study was the identification of new patients with LV systolic dysfunction (defined as a left ventricular ejection fraction of 35% or less) within 30 days of the ED visit. Secondary outcomes were defined as patients with a left ventricular ejection fraction (LVEF) less than 50% found within 30 days of presentation. Both outcomes were determined by ECGs assessed by a deep learning network, an AI-ECG algorithm developed and validated to identify LVEFs of 35% or lower without additional optimization or training.
The median time to ECG after ED visit was 1 day.
In emergency patients with dyspnea, the area under the receiver operating characteristic curve for the AI-ECG algorithm to identify new left ventricular systolic dysfunction was 0.89 (95% CI, 0.86-0.91). The algorithm had an accuracy of 85.9% (95% CI, 84.1-87.6), a specificity of 87%, a sensitivity of 74%, a positive predictive value of 40%, and a negative predictive value of 97%.
The algorithm was also able to identify patients with an LVEF below 50% with an area under the receiver operating characteristic curve of 0.85 (95% CI, 0.83-0.88) with an accuracy of 86% (95% CI, 84.2-87.7). This also achieved a specificity of 91%, a sensitivity of 63%, a positive predictive value of 62%, and a negative predictive value of 92%.
The researchers also evaluated a panel of 866 patients with available N-terminal B-type natriuretic peptide values. NT-proBNP levels greater than 800 pg/mL indicated new LV systolic dysfunction, with an area under the receiver operating characteristic curve of 0.8 (95% CI, 0.76-0.84).
"The current study is retrospective, and prospective studies are needed to assess the impact of AI-ECG on long-term clinical outcomes, which our research team is currently evaluating," Adedinsewo said in an interview.
Adedinsewo added that the technology is currently being used throughout her healthcare system. She told Healio: “This AI-ECG tool is currently available at all Mayo Clinic sites and is accessible through our electronic medical record system, in addition, the tool was recently granted emergency use authorization by the FDA in May for screening confirmed diagnoses or left ventricular dysfunction in patients with suspected novel coronavirus."
The potential to advance patient care
In a related editorial, Dr. Kazi T. Haq of the Knight Cardiovascular Institute at Oregon Health and Science University in Portland, Oregon, and colleagues wrote: "Overall, the findings of Adedinsewo et al. show that— AI using a standard 12-lead ECG lead-wire ECG could improve the identification of new-onset heart failure in emergency department dyspnea patients. This is a strategy that is easy to use in clinical practice and has the potential to significantly improve patient care."