This was a substudy of the prospective multicenter Korean Acute Heart Failure (KorAHF) registry, which enrolled 5625 consecutive patients upon initial hospital admission for acute HF at 10 tertiary university hospitals in Korea. Details on the KorAHF registry objectives, design, and population are available on the clinical trial registration site (ClinicalTrials.gov NCT01389843) and have been published previously [9,10]. Briefly, patients who had signs or symptoms of HF and met one of the following criteria were eligible for enrollment in the KorAHF registry: (1) lung congestion or (2) objective left ventricular systolic dysfunction or structural heart disease findings. There were no exclusion criteria.

In this study, we retrospectively analyzed the prospectively collected data from 1254 patients who were hospitalized for acute HF from March 2011 to February 2014 at 2 out of 10 participating tertiary centers (Seoul National University Bundang Hospital and Severance Hospital) using the KorAHF registry (Figure 1). Additional ECG image data were collected for this study.

This study conformed with the principles outlined in the Declaration of Helsinki. The study protocol was approved by the institutional review board at Seoul National University Bundang Hospital (No. B-1104-125-014) and Severance Hospital (No. 2022-2166-001). The need for written informed consent was waived by the institutional review board. Our research strictly adheres to the Guidelines for Developing and Reporting Machine Learning Predictive Models in Biomedical Research [11].

Data collection methods have been previously described [9]. Briefly, data on patients’ clinical manifestations, biochemical parameters, medication, and outcome were collected using a web-based case report form for up to 60 months by research nurses. Outcome data on patients lost to follow-up were additionally collected from national death records.

The primary endpoint of this study was all-cause mortality. Secondary outcomes included in-hospital outcomes, especially in-hospital mortality. All deaths were considered to be cardiac-related unless a definite noncardiac cause could be established. All outcome data reported from the participating centers were reviewed by an independent clinical event adjudicating committee.

Quantitative ECG (QCG) is an AI analyzer composed of an encoder part and multiple task-specific networks. The encoder part is a modified convolutional neural network with residual connections, squeeze excitation modules, and a nonlocal block. The task-specific networks are multilayer percetron models. The encoder part accepts 2D ECG images as input to produce a common numerical feature vector for downstream tasks. The encoder part was pretrained on 49,731 open ECGs using self-supervised learning schemes and then fine-tuned on 47,194 annotated ECG images of over 32,968 patients who visited the Emergency Department of Seoul National University Bundang Hospital between 2017 and 2019 using multitask learning schemes. The tasks include the classification of 12 rhythms (with 35 subtypes) and production of 10 digital biomarkers correlated with the risk of (1) being critically ill (shock, respiratory failure, or cardiac arrest), (2) cardiac ischemia (acute coronary syndrome, ST-elevation myocardial infarction, or myocardial injury as defined by an elevated troponin level), (3) cardiac dysfunction (pulmonary edema, left and right heart dysfunction, pulmonary hypertension, and clinically significant pericardial effusion), and (4) hyperkalemia. Several validation studies of the system have been published previously [12-14]. The collection of these AI algorithms has been developed into a mobile app (ECG Buddy, ARPI), which has been approved by the Korean Ministry of Food and Drug Safety.

In this study, two QCG features were evaluated: QCG-Critical for critical conditions such as shock or mortality and QCG-HF for a reduced echocardiographic left ventricular ejection fraction (LVEF) of <40%. The QCG scores, representing probability, ranged from 0 to 1.0, with 0 indicating low and 1.0 indicating high probability. With a 9:1 ratio split of the training and test data sets, the internal validation results for these two QCG features showed an area under the curve (AUC) of 0.877 for QCG-Critical and 0.956 for QCG-HF. The composition of the training and validation data sets is presented as a flowchart in Figure 1.

Categorical variables are reported as frequencies (percentages) and continuous variables are expressed as means (SD) or medians (IQR). The two key AI-driven scores (QCG-Critical and QCG-HF) were analyzed as continuous variables. The Student t test and χ² (or Fisher exact) test were used to compare the baseline clinical characteristics between the two groups. The discrimination performance of QCG scores for in-hospital outcomes was evaluated using receiver operating characteristic (ROC) curve analysis. The AUC values were compared using the DeLong test. The logistic regression model was used to estimate the odds ratios (ORs) and 95% CIs. Survival analysis was performed using the Kaplan-Meier method, and the Cox proportional hazard model was used to estimate the hazard ratios (HRs) and 95% CIs for the clinical outcomes. Multivariable analysis was performed with the inclusion of clinically relevant variables.

All tests were two-tailed and a P value <.05 was considered statistically significant. Statistical analyses were performed using R programming version 4.3.0 (The R Foundation for Statistical Computing).

Data of 1254 patients (716 from Seoul National University Bundang Hospital and 538 from Severance Hospital) were analyzed. Among the 1254 patients, 53 (4.2%) experienced in-hospital cardiac death. The baseline characteristics of the study population according to the in-hospital outcomes are shown in Table 1. Compared with survivors, patients who died in the hospital were older, had a higher prevalence of ischemic heart disease, lower LVEF, and higher NT-proBNP levels. By contrast, atrial fibrillation (AF) was more frequent in survivors. The QCG-Critical and QCG-HF scores were significantly higher in patients who experienced in-hospital cardiac death than in survivors (P<.001) (Table 1 and Figure S1 in Multimedia Appendix 1).

In the univariable logistic regression analysis, the QCG-Critical and QCG-HF scores were significant predictors of in-hospital cardiac death (Table 2). Other than QCG scores, echocardiographic LVEF, NT-proBNP level, age, ischemic heart disease, and AF were significantly correlated with in-hospital cardiac death (Table 2).

After adjustment for age, sex, hypertension, diabetes, chronic kidney disease, cerebrovascular disease, ischemic heart disease, valvular heart disease, HF type, AF, and QRS duration, the two QCG scores remained significant predictors of in-hospital cardiac death. Moreover, the QCG-Critical score was an independent predictor of in-hospital cardiac death after further adjustment for echocardiographic LVEF and NT-proBNP level (OR 1.59, 95% CI 1.36-1.87; P<.001).

In a subgroup analysis, the QCG-Critical score was a significant predictor of in-hospital cardiac death regardless of the initial rhythm (AF or sinus rhythm), QRS width (wide or narrow), hypertension, diabetes, HF etiology (ischemic or nonischemic), HF type (de novo or acute decompensated HF), and LVEF (HF with reduced ejection fraction vs HF with preserved or mildly reduced ejection fraction), after adjustment for other clinical parameters (Figure 2).

The QCG-Critical score was significantly higher in patients who experienced cardiac death within 1 day, 2 days, or during hospitalization than in survivors (Figure 3A). When the performance of the QCG-Critical score for predicting these events was analyzed using ROC curves, the AUC values for 1- and 2-day mortality and in-hospital cardiac death were 0.936, 0.917, and 0.821, respectively (Figure 3B).

Comparatively, the AUC values of echocardiographic LVEF and NT-proBNP level for predicting in-hospital cardiac death were 0.642 (P<.001 vs QCG-Critical) and 0.720 (P=.07 vs QCG-Critical) (Figure 4A). The AUC value of the QCG-Critical score (0.821) was significantly (P=.02). higher than that of model 1 (0.705) established using traditional clinical variables, including age, sex, hypertension, diabetes, chronic kidney disease, cerebrovascular disease, ischemic heart disease, valvular heart disease, HF type, AF, and QRS duration. In addition, when the QCG-Critical score was added to model 1, it significantly enhanced the prediction for in-hospital cardiac death (AUC of model 1=0.705 vs AUC of model 1 with QCG-Critical=0.843; P<.001) (Figure 4B). When NT-proBNP and LVEF were further included in model 1 (model 2), the QCG-Critical score again demonstrated additional predictive value for in-hospital cardiac death compared to model 2 alone (AUC of model 2=0.787 vs AUC of model 2 with QCG-Critical=0.863; P=.01) (Figure 4C).

During a median follow-up of 2.7 years, 508 deaths occurred in the study population. To further analyze the performance of the QCG-Critical score for outcome prediction, we divided patients into three QCG-Critical score groups based on arbitrary cut-off values of 0.25 and 0.50 and then conducted survival analysis (Figure 5).

After adjustment for age, sex, comorbidities, HF etiology and type, AF, and QRS widening, patients with higher QCG-Critical scores had significantly higher all-cause mortality rates during follow-up than those with lower QCG-Critical scores (<0.25). The adjusted HRs for patients with QCG-Critical scores between 0.25 and 0.50 and for patients with QCG-Critical scores higher than 0.50 were 1.57 (95% CI 1.28-1.93) and 2.69 (95% CI 2.14-3.38), respectively (all P<.001). With additional adjustment for LVEF and NT-proBNP to the previous model, the adjusted HRs were 1.61 and 2.27, respectively, consistent with the main analysis (Figure S2 in the Multimedia Appendix 1).

In a subgroup analysis, a higher QCG-Critical score (>0.50 vs ≤0.50) was significantly correlated with all-cause mortality during follow-up, regardless of the initial rhythm (AF or sinus rhythm), QRS width (wide or narrow), hypertension, diabetes, HF etiology, HF type (de novo or acute decompensated HF), and LVEF (HF with reduced ejection fraction vs HF with preserved or mildly reduced ejection fraction), after adjustment for other clinical parameters (Figure S3 in Multimedia Appendix 1).