Improving the Accuracy of an Artificial Neural Network Using Multiple Differently Trained Networks

When either detection rate (sensitivity) or false alarm rate (specificity) is optimized in an artificial neural network trained to identify myocardial infarction, the increase in the accuracy of one is always done at the expense of the accuracy of the other. To overcome this loss, two networks that were separately trained on populations of patients with different likelihoods of myocardial infarction were used in concert. One network was trained on clinical pattern sets derived from patients who had a low likelihood of myocardial infarction, while the other was trained on pattern sets derived from patients with a high likelihood of myocardial infarction. Unknown patterns were analyzed by both networks. If the output generated by the network trained on the low risk patients was below an empirically set threshold, this output was chosen as the diagnostic output. If the output was above that threshold, the output of the network trained on the high risk patients was used as the diagnostic output. The dual network correctly identified 39 of the 40 patients who had sustained a myocardial infarction and 301 of 306 patients who did not have a myocardial infarction for a detection rate (sensitivity) and false alarm rate (1-specificity) of 97.50 and 1.63%, respectively. A parallel control experiment using a single network but identical training information correctly identified 39 of 40 patients who had sustained a myocardial infarction and 287 of 306 patients who had not sustained a myocardial infarction (p = 0.003).