759-2 Use of Artificial Neural Networks within Deterministic Logic for Computer ECG Diagnosis of Myocardial Infarction

Abstract

The study was aimed at assessing the effect of incorporating neural networks (NN) inside an existing deterministic computer ECG analysis program in order to enhance the diagnosis of myocardial infarction. Separate neural networks were trained for inferior and anterior myocardial infarction using 200 normals, 100 IMI. 80 AMI and 42 left ventricular hypertrophy cases, all clinically validated. All the networks had a single output to discriminate between MI and non-MI. A variable number of inputs to the networks was used consisting of QRS ± ST-T measurements. Separate test sets including 200 normals. 42 LVH, 101 AMI and 80 IMI cases were then utilised to find the best performing neural networks for IMI and AMI. The best neural network for each of IMI and AMI was then selected and inserted into the existing Glasgow Program (GP) for ECG analysis together with some modifications (M) to the diagnostic logic. The modified and original GP were then assessed using a completely new test set composed of 74 AMI, 52 IMI, 60 LVH and 230 normals. AMI Se IMI Se NSp LSp OSp GP 76% 69% 100% 93% 99% GP+NN+M 78% 88% 100% 85% 97% Se = sensitivity, Sp = specificity, N = normal. L = LVH, Oa = overall This first report of neural networks for the diagnosis of myocardial infarction embedded within a deterministic logic program has shown that (1) the technique significantly improves the diagnosis of inferior though not anterior MI; (2) the evaluation of specificity using only normals is misleading; (3) the technique can usefully be adopted selectively to enhance diagnostic ECG programs in future.