# Artificial Intelligence–Enabled ECG for Diastolic Dysfunction in Congenital Heart Disease: A Novel Risk Stratification Tool

## TL;DR

An AI-enhanced ECG tool accurately assesses diastolic heart function in congenital heart disease patients and predicts mortality risk.

## Contribution

AI-ECG provides a noninvasive, scalable method for diastolic dysfunction grading and mortality prediction in adult congenital heart disease.

## Key findings

- AI-ECG diastolic grades correlate with echocardiographic and invasive hemodynamic measures.
- Higher AI-ECG grades independently predict increased mortality in ACHD patients.
- The model discriminates elevated pulmonary artery wedge pressure with 74% accuracy.

## Abstract

Echocardiography-based assessment of diastolic function in adult congenital heart disease (ACHD) is challenging owing to complex anatomy and heterogenous physiology.

The objectives of the study was to evaluate an artificial intelligence (AI)–enabled electrocardiogram (ECG) (AI-ECG) model for grading diastolic dysfunction in patients with ACHD and assess its correlation with echocardiographic, invasive hemodynamic, and clinical outcomes.

In this single-center retrospective study, we analyzed 6,741 patients from the Mayo Clinic ACHD Registry (median age 37 years, 49% female) followed from 2000 to 2023. The median follow-up was 10 (5-15) years. Using a validated deep neural network (trained on 98,736 ECG-echocardiogram pairs), we assigned an AI-ECG diastolic grade (0-3) to the earliest ECG within 12 months of the index visit. We evaluated associations with echocardiography, hemodynamics, and mortality using nonparametric tests, correlation, Kaplan-Meier curves, Cox regression, and model performance for detecting elevated pulmonary artery wedge pressure (PAWP).

The AI-ECG classified diastolic function as grade 0 in 65.8%, grade 1 in 4.0%, grade 2 in 19.7%, and grade 3 in 10.5%. Higher grades were associated with older age, greater CHD complexity, and more comorbidities including heart failure (6.7% vs 24.8%), diabetes, and cirrhosis (all P < 0.001). N-terminal pro–B-type natriuretic peptide rose with each grade (129 [60-304] to 763 [311-1,915] pg/mL; P < 0.001). AI-ECG pressure estimates correlated with left atrial strain (ρ = −0.52) and right ventricular strain (ρ = −0.50). Invasive hemodynamics followed similar patterns; right atrial pressure rose from 8 (6-11) to 13 (10-18) mm Hg, and wedge pressure from 11 (8-14) to 16 (12-21) mm Hg (P < 0.001). Survival differed by grade (log-rank P < 0.0001); grades 2 and 3 independently predicted mortality (HR: 1.38; 95% CI: 1.09-1.75; HR: 1.63; 95% CI: 1.27-2.08). The model discriminated pulmonary artery wedge pressure ≥20 mm Hg with an area under the receiver operating characteristic curve of 0.74 (95% CI: 0.70-0.78).

AI-ECG diastolic grading correlates with echocardiographic and invasive measures of cardiac filling pressures and independently predicts mortality in ACHD. These findings support its utility as a scalable, noninvasive risk stratification tool.