ECG-AI Classifies LV Dysfunction and HFpEF

Key Takeaways

• Across nearly 1.08 million digital ECGs from 165,243 patients, ECG-based AI was reported to classify reduced EF, midrange EF, and HFpEF from routine tracings.
• Adult discrimination was strongest for reduced EF across both model formats.
• Single-lead performance was close to 12-lead for reduced EF and was more variable for midrange EF and HFpEF across cohorts.
• Clinical data-only models lacked generalizability, and adding clinical variables did not significantly improve ECG-alone performance.

Routine electrocardiography reaches more patients than echocardiographic imaging, yet left ventricular dysfunction and heart failure with preserved ejection fraction often still require imaging before phenotype-specific characterization is clear. That gap has sustained interest in whether one artificial intelligence approach could sort multiple dysfunction patterns from standard and even single-lead tracings, potentially extending signal detection to settings where full imaging is less accessible. Investigators at Wake Forest University tested that question with parallel models built from routine cardiology data.

In the Journal of the American Heart Association, investigators developed two electrocardiogram (ECG)-based artificial intelligence (AI) models, one using 12-lead ECGs and one using single-lead ECGs, to classify left ventricular (LV) dysfunction phenotypes and heart failure with preserved ejection fraction (HFpEF). Reduced ejection fraction (EF) was defined as EF<40 and midrange EF as 40≤mEF<50. Development and validation used nearly 1.08 million digital ECGs from 165,243 patients, and an external pediatric cohort contributed 72,832 ECGs from 42,880 patients.

Among adults, the 12-lead model yielded area under the curve values of 0.90 for reduced EF, 0.81 for midrange EF, and 0.80 for HFpEF. The single-lead model yielded values of 0.89, 0.78, and 0.75 across the same phenotypes. Reduced EF showed the strongest discrimination, and single-lead performance remained close to 12-lead performance for that phenotype.

In the validation cohort, values were 0.92, 0.76, and 0.73 for the 12-lead model and 0.90, 0.75, and 0.74 for the single-lead model across reduced EF, midrange EF, and HFpEF. Clinical data-only machine-learning models lacked generalizability, and adding clinical variables to ECG-based AI did not significantly improve performance over ECG alone. In the pediatric analysis from the external dataset, performance estimates came from a subset of 8,418 ECGs, including 142 cases and 8,276 controls; AUCs were 0.97, 0.71, and 0.64 for the 12-lead model and 0.94, 0.77, and 0.67 for the single-lead model for reduced EF, midrange EF, and HFpEF, respectively, with reduced EF showing the strongest discrimination.

The findings should not be read as establishing HFpEF diagnosis or replacing imaging. Across cohorts, the reported pattern was more favorable for reduced EF than for midrange EF or HFpEF, especially outside the main adult dataset.

The Journal of the American Heart Association study reported that ECG-based AI distinguished multiple LV dysfunction phenotypes from both 12-lead and single-lead tracings, with the clearest performance for reduced EF. The investigators also suggested that single-lead ECG could support potential lower-cost screening, although the available report leaves implementation details and confirmatory imaging requirements unresolved.

Clinician Questions

Which LV dysfunction phenotypes were targeted, and how were reduced EF and midrange EF defined?

The models were reported to classify reduced EF, midrange EF, and HFpEF. Reduced EF was defined as EF<40, and midrange EF as 40≤mEF<50. The available report does not describe the HFpEF labeling criteria or the confirmatory reference standard used for that phenotype.

What remains unclear about using single-lead ECG-AI to screen for HFpEF and other LV dysfunction phenotypes?

The investigators suggested that single-lead ECG could support potential lower-cost screening.