Search

This thesis describes a series of studies on the use of artificial intelligence in computerised electrocardiography, the history of which is first of all reviewed. The relatively recent technique of using artificial neural networks, which aim to simulate the human decision making process, is thereafter introduced. In a preliminary study, the vectorcardiogram derived from the twelve-lead electrocardiogram was studied for the first time in large populations of Caucasians and Chinese. A total of 2058 vectorcardiograms derived from conventional twelve-lead electrocardiograms recorded from 1555 healthy Caucasians and 503 healthy Chinese were analysed to establish the normal limits for Caucasians and Chinese, respectively. Several conclusions can be drawn from this study: (1) Age, sex and race dependent variations are present in the derived vectorcardiogram; (2) In both races, the maximal spatial QRS vector magnitude, as well as the maximal QRS and T vector magnitude in the frontal, horizontal and right sagittal planes, decrease with advancing age in both sexes and are significantly larger in men in all age groups; (3) In groups younger than 40 years, the magnitude of the maximal spatial QRS vector is greater in Caucasians than in Chinese, while in the groups older than 40 years, it is greater in Chinese than in Caucasians; (4) This new data indicates that it is necessary to take racial differences into consideration for interpretation of the derived vectorcardiogram. The first area in which neural networks were studied related to the analysis of cardiac rhythm. In particular, the usefulness of neural networks in separating atrial fibrillation from [sinus rhythm + (supraventricular extrasystoles &/or ventricular extrasystoles)] was investigated. For this purpose, 3080 visually classified ECGs including 2018 with atrial fibrillation and 1062 with [sinus rhythm + (supraventricular extrasystoles &/or ventricular extrasystoles)] were used. Fundamental to this work was the availability of the existing University of Glasgow ECG analysis program which was based on locally developed deterministic logic. This study was divided into five stages: (1) Selection of the optimum parameters for input to a neural network; (2) Determination of the optimum topology of the network; (3) Assessment of the accuracy of the network; (4) Combining the deterministic logic result with the output from the neural network; (5) Modification of the existing logic. Several conclusions can be drawn: (1) A neural network can improve the sensitivity but slightly decreases the specificity in detecting atrial fibrillation compared to the existing deterministic program; (2) Various combinations of the neural network output and the deterministic interpretation are not superior to the use of a neural network alone; (3) Modification of the existing deterministic logic can produce improved performance compared to the use of either a neural network or the original logic. The use of neural networks was also intensively studied in the electrocardiographic diagnosis of myocardial infarction. A total of 1269 electrocardiograms [515 from patients with clinically documented myocardial infarction (255 anterior and 260 inferior), 144 from patients with clinically validated left ventricular hypertrophy and 610 from normals] were used to study the usefulness of the neural network approach. This study comprised a series of six experiments. Experiments 1, 2, 3 and 4 concerned the use of neural networks in isolation, while experiments 5 and 6 studied the effects of implanting neural networks into the deterministic program. The conclusions drawn from experiments 1-4 on the use of neural networks in isolation for the diagnosis of myocardial infarction are that: (1) There is no significant benefit from using derived vectorcardiographic measurements in addition to electrocardiographic QRS and ST-T parameters as input variables to the neural network; (2) In the two-group situation (normal versus myocardial infarction), the neural network is superior to deterministic logic for the detection of myocardial infarction; (3) In the diagnosis of anterior myocardial infarction, neural networks trained in the three-group situation (normal versus myocardial infarction versus left ventricular hypertrophy) perform better than those trained in the two-group situation, whereas in the diagnosis of inferior myocardial infarction, there is no benefit from training in the three-group situation; (4) The alteration of the network topology to include three output neurons rather than one does not lead to any improvement in the diagnosis of myocardial infarction; (5) Neural networks using electrocardiographic QRS and ST-T measurements as input variables are superior to the neural networks using QRS measurements only; (6) The enhanced sensitivity of detecting myocardial infarction by the neural network is more pronounced in inferior myocardial infarction than in anterior myocardial infarction; (7) The specificity of diagnosing myocardial infarction by the neural network in the left ventricular hypertrophy cases decreases compared to the original logic no matter what training methodology is used.