Improving clinical decision support using data mining techniques

Physicians, in their ever-demanding jobs, are looking to decision support systems for aid in clinical diagnosis. However, clinical decision support systems need to be of sufficiently high accuracy that they help, rather than hinder, the physician in his diagnosis. Decision support systems with accuracies, of patient state determination, of greater than 80% are generally perceived to be sufficiently accurate to fulfil this role of helping the physician. We have previously shown that Data Mining techniques have the potential to provide the underpinning technology for clinical decision support systems'. Tn this paper, an extension of the work in reference 2., we describe how changes in Data Mining methodologies, for the analysis of 12-lead ECG data improve the accuracy by which Data Mining algorithms determine which patients are suffering from heart disease. We show that the accuracy of patient state prediction, for all the algorithms which we investigated, can be increased by up to 6% using the combination of appropriate test training ratios and 5 fold cross-validation. The use of cross-validations greater than 5 fold appears to reduce the improvement in algorithm classification accuracy gained by the use of this validation method. The accuracy of 84% in patient state predictions obtained using the algorithm OC1 suggests that this algorithm will be capable of providing the required accuracy for clinical decision support systems. Further work is in progress, in conjunction with Odense University Hospital, which is investigating the design and implementation of a multi-domain clinical decision support system underpinned by Data Mining.