AI-ECG for One-Year Mortality Risk Prediction

Overview

Cardiovascular disease (CVD) remains one of the leading causes of death worldwide. While the electrocardiogram (ECG) is a standard, widely accessible tool for cardiovascular screening, traditional risk assessment models often rely heavily on blood test results, which may be unavailable in electronic health records (EHRs). To address this limitation, the Chang Gung ECG Mortality Risk Prediction Software, an artificial intelligence (AI)-based Software as a Medical Device (SaMD), was developed. The software analyzes standard 10-second, 12-lead resting ECG signals to predict the probability of cardiac-related mortality within one year. This study is a multicenter retrospective cohort study designed to validate the clinical performance of the AI software. Researchers will analyze retrospectively collected ECG data from patients aged 20 years or older with suspected cardiovascular disease across three hospitals in Taiwan. The AI model's predictions will be compared with the actual one-year mortality outcomes documented in the patients' medical records. The primary objective is to determine whether the AI model can accurately and consistently stratify patients according to their risk of cardiac-related mortality (e.g., heart failure, arrhythmia, and myocardial infarction), with an area under the receiver operating characteristic curve (AUC) greater than 0.80. The software is intended to serve as a clinical decision-support tool for long-term risk stratification in non-acute clinical settings, thereby assisting physicians in clinical decision-making and long-term patient management.

Cardiovascular disease (CVD) is a major global health burden. Current cardiovascular risk assessment models (e.g., Framingham and QRISK) rely heavily on blood biochemical data, which limits their applicability when electronic health record (EHR) data are incomplete. The 12-lead resting electrocardiogram (ECG) is a rapid, non-invasive, and widely accessible screening tool. Recent advances in artificial intelligence (AI), particularly deep learning models such as ResNet, have demonstrated superior capabilities for automatically extracting clinically relevant features from ECG signals to predict cardiovascular risk. This national multi-center retrospective study aims to evaluate the clinical performance of the Chang Gung ECG Mortality Risk Prediction Software, a standalone Software as a Medical Device (SaMD). The core algorithm utilizes a 1D-ResNet-18 convolutional neural network to analyze 10-second, 12-lead resting ECG signals sampled at 500 Hz with a 60-Hz Alternating Current (AC) filter. The software outputs the predicted probability of cardiac-related mortality within one year to assist physicians in non-acute clinical settings. Study Methodology: The study will retrospectively collect de-identified electronic health records and ECG data obtained between August 2011 and September 2024 from three institutions in Taiwan: Tri-Service General Hospital, Kaohsiung Armed Forces General Hospital, and Taipei Municipal Wanfang Hospital. Only the first eligible ECG from each patient will be included to avoid intra-individual bias. The AI model's predictions will be compared with the actual one-year mortality outcomes. To improve interpretability, cardiologists with more than five years of clinical experience will review high-risk predictions using Gradient-weighted Class Activation Mapping (Grad-CAM). A prediction will be considered clinically interpretable only if both the predicted risk and the corresponding Grad-CAM localization are deemed clinically reasonable. Statistical Analysis: The study employs a one-tailed superiority design with a significance level of 0.05. The null hypothesis states that the area under the receiver operating characteristic curve (AUC) is ≤0.80, whereas the alternative hypothesis states that the AUC is \>0.80. Subgroup analyses will be performed according to age groups (e.g., 20-40, 41-60, and \>60 years) and disease categories (e.g., arrhythmia, myocardial infarction, and heart failure). DeLong's test will be used to evaluate the model's predictive performance across different demographic and clinical subgroups. Data Privacy and Federated Learning: All patient data will be strictly de-identified in accordance with Health Insurance Accountability and Portability Act (HIPAA) guidelines and analyzed within secure, closed intra-hospital networks. If the initial validation does not achieve the predefined performance target (AUC \>0.80), a horizontal federated learning architecture will be implemented. Up to 10% of the available dataset will be used for model fine-tuning, after which the updated model will be independently validated using a separate untouched dataset to ensure robustness and prevent data contamination.