Abstract 2106: A Resting ECG Neural Network Predicts Cardiovascular Mortality
Background: The resting 12-lead ECG remains one of the most commonly used studies when evaluating patients with suspected cardiovascular disease. Although the prognostic values of individual findings of the ECG have been calculated, a neural network that takes into account multiple ECG characteristics could be used to more accurately predict cardiovascular mortality.
Objective: To create an artificial neural network that can use resting ECG characteristics and basic demographic information to predict cardiovascular mortality.
Methods: 45,855 resting ECGs ordered by physician’s discretion for usual clinical indications between 1987 and 2000 in a Veteran’s hospital were available for analysis. Age, gender, race, and BMI were recorded and the population was followed until 2002 using the California Death Index. Baseline ECG characteristics were recorded and analyzed using the GE/Marquette computerized ECG system. 41,269 (90%) ECGs were used to train artificial neural networks to predict cardiovascular mortality based on 132 ECG characteristics and the four demographic variables. The ability of the networks to predict cardiovascular mortality was then tested in the remaining 4,586 (10%) ECGs.
Results: Over a mean follow up of 7.5 years, there were a total of 449 (9.8%) cardiovascular deaths in the test group. Using a back-propagation algorithm, a neural network consisting of one hidden layer with 70 nodes was trained to produce an output score between 0 and 1 that was predictive of mortality. When tested, the ROC curve generated by this network had an AUC of 0.78 (95% CI: 0.76 - 0.80).
Conclusion: By creating an artificial neural network that uses multiple ECG findings and a few demographic characteristics, a computerized prediction of cardiovascular mortality can be made.