Prediction of mortality from 12-lead electrocardiogram voltage data using a deep neural network.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Electrocardiograms (ECGs) are standard for cardiac monitoring.
• Predicting future clinical events from ECGs remains an area for advancement.

Purpose of the Study:

• To investigate if a deep neural network (DNN) can predict 1-year all-cause mortality from ECG data.
• To assess the prognostic value of DNN-analyzed ECGs, particularly in cases deemed normal by clinicians.

Main Methods:

• Trained a DNN on 1,169,662 12-lead resting ECGs from 253,397 patients over 34 years.
• Validated the model on a separate test set of 168,914 patients.
• Compared DNN predictions with physician interpretations and expert cardiologist reviews.

Main Results:

• The DNN achieved an AUC of 0.88 for predicting 1-year mortality.
• Performance remained high (AUC=0.85) even for ECGs interpreted as 'normal' by physicians.
• Features identified by the DNN were often not apparent to expert reviewers.

Conclusions:

• Deep learning significantly enhances the prognostic information derived from 12-lead resting ECGs.
• DNNs can identify mortality risk in ECGs that appear normal to human experts.
• This technology offers a powerful tool for risk stratification and clinical decision-making.