[Paper Review] Deep neural networks can predict mortality from 12-lead electrocardiogram voltage data

Sushravya Raghunath, Alvaro Ulloa+13 more|arXiv (Cornell University)|Apr 15, 2019

ECG Monitoring and Analysis45 references50 citations

TL;DR

A deep neural network predicts one-year all-cause mortality from 12-lead resting ECG voltage-time traces, achieving high AUC even in ECGs labeled as normal by physicians.

ABSTRACT

The electrocardiogram (ECG) is a widely-used medical test, typically consisting of 12 voltage versus time traces collected from surface recordings over the heart. Here we hypothesize that a deep neural network can predict an important future clinical event (one-year all-cause mortality) from ECG voltage-time traces. We show good performance for predicting one-year mortality with an average AUC of 0.85 from a model cross-validated on 1,775,926 12-lead resting ECGs, that were collected over a 34-year period in a large regional health system. Even within the large subset of ECGs interpreted as 'normal' by a physician (n=297,548), the model performance to predict one-year mortality remained high (AUC=0.84), and Cox Proportional Hazard model revealed a hazard ratio of 6.6 (p<0.005) for the two predicted groups (dead vs alive one year after ECG) over a 30-year follow-up period. A blinded survey of three cardiologists suggested that the patterns captured by the model were generally not visually apparent to cardiologists even after being shown 240 paired examples of labeled true positives (dead) and true negatives (alive). In summary, deep learning can add significant prognostic information to the interpretation of 12-lead resting ECGs, even in cases that are interpreted as 'normal' by physicians.

Motivation & Objective

• Motivate using ECG voltage-time traces to forecast future mortality risk.
• Assess whether a deep learning model can predict one-year all-cause mortality from resting 12-lead ECGs.
• Evaluate model performance on a large, long-span real-world dataset.
• Investigate whether prognostic signals exist even in ECGs interpreted as normal by physicians.

Proposed method

• Train a deep neural network on 1,775,926 resting 12-lead ECGs collected over 34 years.
• Evaluate predictive performance using AUC for one-year mortality.
• Assess performance within the subset of ECGs labeled as normal by physicians.
• Use a Cox Proportional Hazards analysis to relate predicted groups to long-term mortality (hazard ratio).
• Conduct a blinded survey with three cardiologists to compare model-recognized patterns to human interpretation.

Experimental results

Research questions

• RQ1 Can a deep neural network predict one-year mortality from 12-lead resting ECG voltage-time traces?
• RQ2 Does predictive performance persist within ECGs deemed normal by physicians?
• RQ3 What is the association between model-predicted mortality and long-term outcomes (hazard ratio) in follow-up?
• RQ4 Are the patterns learned by the model visually apparent to clinicians?

Key findings

• Average AUC of 0.85 for predicting one-year mortality on cross-validated data.
• Within the normal-ECG subset (n=297,548), AUC remains high at 0.84.
• Cox model hazard ratio of 6.6 (p<0.005) for dead vs alive at one year over a 30-year follow-up.
• Model patterns were generally not visually apparent to cardiologists after viewing paired examples.
• Deep learning adds significant prognostic information to resting 12-lead ECG interpretation, even when labeled normal.