Skip to Main Content
In this study, a SPECT dataset containing the records of 267 patients with a variety of heart diseases was analysed using a combined rough sets and neural network approach. The dataset consists of a collection of binary features representing thresholded intensity levels (perfusion levels measured via 201Tl radiography) from SPECT images of the heart, taken at rest and after stress. It should be noted that the subjects with a variety of heart diseases were collectively grouped under the single category of `abnormal'. The rough sets data mining approach provides a human readable rule based classification scheme that is used to classify the objects in the database. The results from this preliminary study yielded an overall classification accuracy of over 93% (± 3.8), with a specificity and sensitivity of 85% and 95% respectively. Further, the rough sets analysis was able to reduce the feature space from 22 to 10 essential features, with a minimal reduction in classification accuracy. The reduced feature set was tested for classification accuracy independently using a neural network approach. The resulting classification accuracy was significantly lower than the obtained from a rough sets based approach. This result is significant in that it provides further evidence that small biomedical datasets are very noisy, with a significant number of local minima that render them resistant to standard machine learning approaches. One must be prepared to try alternative approaches, such as rough sets, that do not rely on smooth energy landscapes.