AI-enabled ECG index for predicting left ventricular dysfunction in patients with ST-segment elevation myocardial infarction

Study design and populationThis retrospective study was designed to validate the prognostic value of an AI-enabled ECG derived probability index as a digital biomarker for predicting the prognosis of patients with STEMI after primary PCI. We retrospectively included consecutive patients who underwent primary PCI for STEMI between January 2019 and June 2022 in our institution. STEMI was diagnosed by a new ST segment elevation in ≥ 2 contiguous leads, measuring > 0.2 mV in leads V1-3 or 0.1 mV in all other leads on a 12-lead ECG in a patient with the acute onset of chest pain or dyspnea. Patients who died within 24 h, those who underwent coronary artery bypass grafting, and patients for whom adequate serial ECGs could not be obtained were excluded from the study. This study was approved by the Institutional Review Board (IRB) of Biomedical Research Institute in Seoul National University.Bundang Hospital (B-2201–733-104) and all methods were performed in accordance with the relevant guideline and regulation. The requirement for informed consent was waived by the IRB because of the retrospective study design using fully anonymized ECG and health data, with minimal potential harm.Development of an AI-enhanced ECG prediction model for STEMIIn previous research, we developed a deep-learning algorithm for the detection of AMI and STEMI using a 12-lead ECG10,27. The performance of this algorithm was validated, substantiating its efficacy. The output values, situated between 0 and 1, denote that an elevated value corresponds to an increased probability of the presence of AMI or STEMI. Data were sourced from two different hospitals for internal and external validation. A total of 22,259 ECGs (13,916 ECGs of non-AMI and 8,343 ECGs of AMI) from 15,113 patients were used to develop and validate the model. The area under the receiver operating characteristics curve (AUROC) was 0.927 (95% confidence interval [CI]: 0.908–0.945) for AMI and 0.983 (95% CI: 0.976–0.989) for STEMI.We developed a ResNet-based model using PyTorch and Python, which is typically applied in image recognition, to classify ECG data by recognizing complex patterns. ECG classification is challenging because it involves deciphering rhythm and morphological features, both in image and time series analysis. Figure 5 shows the architecture of our deep-learning model, which includes a stem block, followed by six residual blocks, and ends with a fully connected layer to discern these patterns. Features from the ECG are progressively extracted through each block. We grouped every two residual blocks into what we call a “stage”; our model comprises three such stages. Within each residual block, we arranged a sequence of layer: a one dimensional (1D) convolutional layer, batch normalization, ReLU activation, another 1D convolutional layer, additional batch normalization, a dropout layer, and a skip connection. The stem block has a single layer plus a skip connection and max pooling. For training, we chose the Adam optimizer paired with a cosine warm-up optimization scheduler to adjust the weights and focal loss function for learning. We ran this scheduler for 150 epochs to test different hyperparameters and determine the best model structure, focusing on the highest area under the receiver operating characteristics (AUROC).Figure 5Architecture of the deep learning model. We utilized a ResNet-based model to predict AMI and STEMI. It consists of three parts: a stem block, stage block, and a full connected layer. AMI, acute myocardial infarction; STEMI, ST elevated myocardial infarction.Data collection and analysis of serial ECGsSerial ECGs were analyzed at multiple time points to quantify the dynamic ECG change and to evaluate myocardial reperfusion following primary PCI in patients with STEMI. The AI-enabled ECG quantified the probability of STEMI at each time points as a continuous variable between 0 and 1, i.e., as a digital biomarker. The time points at which the ECG was taken were as follows (Supplementary Fig. S4):

More than 1 month prior to the STEMI event (Baseline)

Time of arrival to the emergency department prior to PCI (Pre-PCI)

Immediately after primary PCI (Immediate Post-PCI)

Six hours after PCI (6 h Post-PCI)

Twenty-four hours after PCI (24 h Post-PCI)

Time of discharge (At discharge)

One month following the PCI (1 M Post-PCI)

Echocardiographic assessments were usually performed the day after primary PCI and follow-up echocardiography was performed around 6 months after the onset of STEMI. LV end systolic volume, LV end systolic volume, and LV ejection fraction (LVEF) were estimated using the bi-planar modified Simpson’s rule from apical two and four chamber views.Study endpointThe primary endpoint was post-STEMI LV dysfunction, defined as a LVEF < 50% during hospitalization and at the 6-month follow-up. The secondary endpoint was cardiac death or heart failure hospitalization within 1 year of the primary PCI. Cardiac death was defined as a mortality with a defined cardiovascular cause and heart failure hospitalization was defined as admission for ≥  24 h with a primary diagnosis of heart failure, with ≥ 1 symptom and ≥ 2 physical examination, laboratory, or invasive findings of heart failure, and receives a heart failure-specific treatment28.Statistical analysisContinuous variables are presented as mean (± standard deviation) and were compared using the unpaired and paired Student’s t-test or Mann–Whitney U test. Means of tertile groups were compared by analysis of variance (ANOVA). Categorical variables are expressed as frequencies and percentages and were compared using the Pearson χ2 test or Fisher’s exact test when the Cochran rule was not met. For the analysis of probability index values obtained at different time points, we used the median value and the statistical significance of differences in these median values across various time points was evaluated using the non-parametric Kruskal–Wallis H-test. Univariate and multivariate logistic regression analyses were performed to identify the proportional hazard risk for LV dysfunction in patients with STEMI after primary PCI, which were adjusted for known potential confounders (age, sex, body mass index, systolic blood pressure, diastolic blood pressure, pulse rate, Killip class ≥ 3, hypertension, diabetes mellitus, end stage renal disease, multiple vessel coronary artery disease, and infarct-related artery). Statistical significance was set at P-value < 0.05. All statistical analyses were performed using IBM/SPSSv24.0 (IBM/SPSS, Chicago, IL, USA), RStudio (Integrated Development Environment for R. RStudio, PBC, Boston, MA, USA), and Python (Python Software Foundation, Wilmington, DE, USA).