ECG-surv: A deep learning-based model to predict time to 1-year mortality from 12-lead electrocardiogram

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Informatics

Background:

• Electrocardiogram (ECG) abnormalities show prognostic value for patient survival.
• Traditional statistical models struggle to fully utilize complex, unstructured ECG data for survival prediction.

Purpose of the Study:

• To introduce and evaluate a deep-learning model, ECG-surv, for survival analysis using both censored and unstructured ECG data.
• To predict 1-year mortality by extracting unique features from 12-lead ECG data.

Main Methods:

• Developed ECG-surv, a deep neural network comprising feature extraction and time-to-event analysis components.
• Evaluated ECG-surv on independent and external test sets using different ECG devices.
• Compared ECG-surv against Cox proportional models and the Framingham risk Cox model.

Main Results:

• ECG-surv significantly outperformed the Cox model in predicting 1-year all-cause mortality (C-index 0.860 vs. 0.796 on test set).
• ECG-surv demonstrated superior prediction for cardiovascular death (C-index 0.891) compared to the Framingham risk Cox model (C-index 0.734).
• The model showed high predictive accuracy on both independent and external datasets.

Conclusions:

• ECG-surv effectively leverages unstructured ECG data for survival analysis.
• The deep-learning model surpasses traditional statistical methods in predicting patient mortality and cardiovascular events.
• ECG-surv represents a valuable advancement for patient survival prediction.