Abstract Body (Do not enter title and authors here): Introduction: Recent research has demonstrated the ability of artificial intelligence (AI) to categorize patients with reduced left ventricular ejection fraction (LVEF) using 12-lead electrocardiograms (ECGs). However, there remains a gap in research concerning the continuous prediction of LVEF values, despite its potential clinical significance.
Hypothesis: We hypothesized that continuous prediction of LVEF values is achievable using 12-lead electrocardiograms.
Aims: Our aim was to develop predictive models for LVEF, both as a continuous outcome and as a binary outcome using a defined cutoff value of 40%.
Methods: This retrospective, cross-sectional, two-center study comprised 21,471 patients who underwent echocardiography and ECG within a 2-week period. One center contributed data for model derivation (14,203 patients), while the other served for external validation (7,268 patients). We employed four distinct machine learning methodologies encompassing regression and classification tasks, each with unique conceptual frameworks. Predictive variables included age, gender, and 10 parameters from ECGs, commonly measurable automatically with standard 12-lead ECGs. Additionally, conditional branching was performed based on the presence of atrial fibrillation (AF) during ECG recording due to the absence of P-axis and P-R interval in AF patients.
Results: In regression analyses, the R-squared values for all four machine learning methodologies were less than 0.75 for both non-AF and AF patients in internal testing, and even lower in external validation (each R-squared < 0.25). Conversely, for LVEF classification in internal validation, area under curves (AUCs) for detecting LVEF<40% ranged from 0.93 to 1.0 in non-AF patients and 0.88 to 1.0 in AF patients. In external validation, these AUCs ranged from 0.78 to 0.90 in non-AF patients and 0.72 to 0.81 in AF patients.
Conclusions: Consistent with prior studies, the classification of reduced LVEF using electrocardiograms demonstrates high predictive value, especially in the absence of AF. However, continuous value estimation through regression proves challenging, revealing inherent limitations in LVEF prediction solely from electrocardiographic data.