AI-ECG for One-Year Mortality Risk Prediction

Cardiovascular disease (CVD) is a major global health burden. Current risk assessment models (e.g., Framingham, QRISK) rely heavily on blood biochemistry data, which limits their applicability when electronic health record (EHR) data is incomplete. The 12-lead resting electrocardiogram (ECG) is a rapid, non-invasive, and highly accessible screening tool. Recent advancements in artificial intelligence (AI), specifically deep learning networks (such as ResNet), have demonstrated superior automatic feature extraction capabilities from ECG signals for predicting CVD risks.

This national multi-center retrospective study aims to evaluate the efficacy of a standalone Medical Device Software (SaMD), the Chang Gung ECG Mortality Risk Prediction Software. The core algorithm utilizes a 1D-ResNet-18 convolutional neural network to analyze 10-second, 12-lead resting ECG digital signals (sampled at 500Hz with a 60Hz AC filter). The software outputs a one-year mortality risk probability related to cardiac conditions to assist physicians in non-acute clinical settings.

Study Methodology The study will retrospectively collect and de-identify electronic health records and ECG data (from August 2011 to September 2024) across three institutions in Taiwan: Tri-Service General Hospital, Kaohsiung Armed Forces General Hospital, and Taipei Municipal Wanfang Hospital. Only the first eligible ECG per patient is included to prevent intra-individual bias.

The AI model's predictions will be compared against the actual one-year mortality outcomes. To ensure interpretability, cardiologists with over 5 years of clinical experience will review high-risk predictions using Gradient-weighted Class Activation Mapping (Grad-CAM). A prediction is considered correct only if both the risk assessment and the Grad-CAM localization are 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, while the alternative hypothesis targets an AUC > 0.80. Subgroup analyses will be conducted based on age distributions (e.g., 20-40, 41-60, >60 years) and specific cardiac etiologies (e.g., arrhythmias, myocardial infarction, heart failure) using DeLong's test to evaluate the model's predictive performance across different demographic and clinical scenarios.

Data Privacy and Federated Learning All patient data will be strictly de-identified according to HIPAA guidelines and analyzed within closed, secure intra-hospital networks. If the initial validation fails to meet the expected performance (AUC > 0.80), a federated learning approach will be initiated using a horizontal architecture. A maximum of 10% of the dataset will be used for model fine-tuning, after which the updated model will be independently re-validated using untouched data to ensure robustness and prevent data contamination.