Validation of electrocardiographic artificial intelligence model for outcome prediction in patients with heart failure

Abstract Introduction Although several valuable biomarkers have been introduced for heart failure (HF) patients, their utilization in routine clinical practice is often constrained by cost and limited availability. Electrocardiogram (ECG) is an essential and cost-effective tool for evaluating cardiovascular diseases. We tested artificial intelligence (AI) algorithm analyzing a printed ECG image for the outcome prediction in patients with acute HF. Methods Two tertiary centers prospectively enrolled 1,254 patients with acute HF syndrome. Baseline ECG images were analyzed using a deep learning system called Quantitative ECG (QCG™), which features a CNN-based binary classifier trained to detect several urgent clinical conditions including shock, cardiac arrest, myocardial infarction, and HF, in addition to rhythm diagnosis. Results Patients with reduced left ventricular ejection fraction (LVEF, <40%) had significantly higher QCG scores for LV dysfunction (QCG-LVdys) (0.80 ± 0.20 vs. 0.38 ± 0.27, P < 0.001), and the AUC of QCG-LVdys for low LVEF was 0.884. Fifty-three patients (4.2%) resulted in in-hospital cardiac death (IHCD), and QCG score for critical events (QCG- Critical) was significantly higher in these patients than in survivors (0.57 ± 0.23 vs. 0.29 ± 0.20, P < 0.001). QCG-Critical was an independent predictor for IHCD after adjustment for age, sex, comorbidities, etiology/type of HF, atrial fibrillation, and QRS widening (adjusted OR = 1.68 [95% CI, 1.47 – 1.92], P <0.001, per 0.1). After further adjustments for echocardiographic LVEF and NTproBNP, QCG-Critical was still significantly correlated with IHCD (adjusted OR = 1.59 [95% CI, 1.36 – 1.87], P <0.001, per 0.1). The AUC of QCG-Critical for IHCD was 0.821 which was higher than the AUC of echocardiographic LVEF (P<0.001) or NTproBNP (P=0.07) (Fig 1A). The AUC of the multivariate model with QCG-Critical and other covariates was 0.866 (Fig 1B). During long-term follow-up, patients with higher QCG-Critical scores (>0.5) showed higher all-cause mortality rate compared to those with low QCG-Critical scores (<0.25) (adjusted HR = 2.69 [2.14 – 3.38], P < 0.001) (Fig 2). Conclusion Predicting outcomes in patients with acute HF appeared feasible using the newly developed Quantitative ECG (QCG) scores. These results suggest the possibility that QCG may serve as a novel biomarker for HF patients.ROC curves for In-hospital cardiac deathSurvival curves according to QCG scores