Phenotype specific performance of an artificial intelligence enhanced ECG model for detecting left ventricular systolic dysfunction across the heart failure spectrum

Background

Heart failure (HF) encompasses a heterogeneous spectrum of phenotypes defined by left ventricular ejection fraction (LVEF). Artificial intelligence–enhanced electrocardiography (AI-ECG) has emerged as a scalable approach for identifying left ventricular systolic dysfunction (LVSD), but its performance across HF phenotypes remains incompletely characterized.

Objectives

To evaluate the diagnostic performance of an AI-based ECG model for detecting LVSD across heart failure phenotypes defined by cardiac magnetic resonance (CMR).

Methods

This retrospective observational study included 384 patients who underwent ECG, transthoracic echocardiography, and CMR within a 30-day interval. ECGs were analyzed using the PMcardio® AI-ECG LVSD detection algorithm, which classifies left ventricular function as preserved (≥50%), mildly reduced (41–49%), or reduced (≤40%). CMR served as the reference standard. Diagnostic performance metrics, including sensitivity, specificity, predictive values, accuracy, and area under the receiver operating characteristic curve (AUC), were assessed and stratified by CMR-defined HF phenotype: HF with preserved ejection fraction (HFpEF), mildly reduced ejection fraction (HFmrEF), and reduced ejection fraction (HFrEF).

Results

Left ventricular systolic dysfunction (LVEF <50%) was present in 112 patients (29.2%). According to CMR, 272 patients were classified as HF with preserved ejection fraction (HFpEF), 24 as HF with mildly reduced ejection fraction (HFmrEF), and 88 as HF with reduced ejection fraction (HFrEF).

Overall, the PMcardio® AI-ECG LVSD detection algorithm demonstrated good diagnostic performance for identifying LV systolic dysfunction, with an AUC of 0.86. Diagnostic performance varied across HF phenotypes. Sensitivity was highest in patients with HFrEF (92.0%), intermediate in HFmrEF (75.0%), and lower in HFpEF (69.9%). In contrast, specificity was highest in HFpEF (89.0%), compared with HFmrEF (83.0%) and HFrEF (78.4%). Negative predictive value exceeded 90% in HFpEF, supporting the role of AI-ECG as an effective rule-out tool in patients with preserved systolic function.

This phenotype-dependent gradient highlights improved detection of advanced systolic dysfunction, with decreasing sensitivity across milder phenotypes.

Conclusions

The diagnostic performance of the PMcardio® AI-ECG LVSD detection algorithm varies across the heart failure spectrum, with optimal sensitivity in HFrEF and high specificity in HFpEF. These findings support the use of AI-enhanced ECG as a scalable screening tool tailored to the heterogeneous phenotypes of heart failure.Performance across the sprectrum.For image description, please refer to the figure legend and surrounding text.