Feature and classifier fusion for 12-lead ECG classification

Two methodologies, feature and classifier fusion, for the problem of computerized 12-lead electrocardiogram classification, are investigated. Firstly, the entire classification problem is subdivided into a number of smaller bi-dimensional ones. By employing bi-group Neural Network classifiers, independent feature vectors for each diagnostic class are examined individually and the output from each classifier are fused together to produce one single result. Secondly, two classifiers, namely the aforementioned and a decision tree, are fused together through a novel approach of a Specificity Matrix. This methodology addresses the problem of unresolved conflict during fusion of classifiers and aims to exploit the merits of each classifier and suppress their weaknesses. 290 validated 12-lead electrocardiogram recordings, comprising six diagnostic classes, were used to train, validate and test both methodologies. The framework of bi-group classifiers enhanced the overall performance by 12.0% in comparison with conventional approaches. In the second instance, the fusion of the two classifiers produced a performance level of 81.3%; superior to either classifier in isolation. This approach offers a viable solution to the unresolved problem of conflict between classifiers during fusion and can be extended readily to accommodate any number of diagnostic classes and classifiers.