Analysis of Electrocardiograms Using Artificial Neural Networks.

Most conventional ECG interpretation programs use decision tree logic for interpretation of the ECG. The performance is generally good but can be improved. Artificial neural networks represent a new computer method, which has proved to be of value in pattern recognition and classification tasks.

The purpose of the studies in this thesis was to improve the analysis/interpretation of the 12-lead ECG by using artificial neural networks. The input values to the networks are extracted from the measurement section of a commercially available interpretation program. No special recording technique or devices have to be used.

The results show that artificial neural networks improve computerized ECG interpretation for the diagnosis of acute and healed myocardial infarction. They also perform well in quality control of the ECG recordings by detecting lead reversals with high sensitivity and specificity.

The output values from an accurately trained neural network can, under certain conditions, be regarded as a posteriori probabilities for a diagnosis. The output values can also be transformed to verbal statements concerning different probability levels for healed myocardial infarction. The agreement between these probability estimates and those of an experienced electrocardiographer was high.

The results indicate that artificial neural networks, if properly trained and validated, will be a useful aid in the attempt to improve the diagnostic yield of the 12-lead ECG.