Artificial intelligence-enabled electrocardiography for identifying heart failure in patients with dyspnea in the emergency department

Background

Not all patients presenting to the emergency department with dyspnea have acute heart failure (AHF). N-terminal pro–B-type natriuretic peptide (NT-proBNP) is considered gold standard to rule out AHF in this population; however, its use is associated with increased cost and potential delays in diagnosis due to laboratory turnaround time.

Purpose

To evaluate whether an artificial intelligence–enabled electrocardiography (AI-ECG) marker can identify AHF in patients presenting with dyspnea and to compare its diagnostic performance with that of NT-proBNP.

Methods

A total of 4,373 consecutive patients presenting to the emergency department with dyspnea between January 2020 and June 2022 at a single tertiary hospital were retrospectively evaluated. An AI-ECG model was trained using an independent dataset collected between 2009 and 2019 and subsequently applied to the present cohort. The model generated AI-ECG–derived features for pulmonary edema (ai-PulEdema), left ventricular dysfunction (ai-LVDys), and natriuretic peptide–like signals (ai-NP). AHF was defined as hospitalization with heart failure as the primary diagnosis and treatment with intravenous loop diuretics.

Results

Overall, 617 patients (14.1%) were diagnosed with AHF. Compared with patients without AHF, those with AHF were older, more frequently male, had higher NT-proBNP levels, and lower C-reactive protein levels. AI-ECG feature scores were significantly higher in patients with AHF than in those without: ai-PulEdema (70.6 ± 24.2 vs. 32.0 ± 29.1, P<0.001), ai-LVDys (51.0 ± 33.9 vs. 19.9 ± 25.8, P<0.001), and ai-NP (35.5 ± 15.0 vs. 13.1 ± 12.1, P<0.001). Diagnostic performance for identifying AHF was high for ai-PulEdema (AUC 0.83; 95% CI 0.81–0.84) and ai-NP (AUC 0.81; 95% CI 0.80–0.83), comparable to that of NT-proBNP (AUC 0.81; 95% CI 0.79–0.82).

In the subgroup of patients with advanced chronic kidney disease (serum creatinine >3.0 mg/dL), NT-proBNP showed attenuated discrimination (AUC 0.67, 95% CI 0.61-0.73), whereas ai-PulEdema maintained preserved diagnostic performance (AUC 0.73, 95%vCI 0.66-0.80).

Conclusions

Among patients presenting to the emergency department with dyspnea, AI-ECG–derived parameters demonstrated diagnostic performance comparable to NT-proBNP for identifying AHF. These findings suggest that AI-ECG may facilitate early identification of AHF in this population.Prediction of AHFFor image description, please refer to the figure legend and surrounding text.Subgroups according to sex and CKDFor image description, please refer to the figure legend and surrounding text.