Predicting cardiovascular risks using pattern recognition and data mining

This thesis presents the use of pattern recognition and data mining techniques into risk prediction models in the clinical domain of cardiovascular medicine. The data is modelled and classified by using a number of alternative pattern recognition and data mining techniques in both supervised and unsupervised learning methods. Specific investigated techniques include multilayer perceptrons, radial basis functions, and support vector machines for supervised classification, and self organizing maps, KMIX and WKMIX algorithms for unsupervised clustering. The Physiological and Operative Severity Score for enUmeration of Mortality and morbidity (POSSUM), and Portsmouth POSSUM (PPOSSUM) are introduced as the risk scoring systems used in British surgery, which provide a tool for predicting risk adjustment and comparative audit. These systems could not detect all possible interactions between predictor variables whereas these may be possible through the use of pattern recognition techniques. The thesis presents KMIX and WKMIX as an improvement of the K-means algorithm; both use Euclidean and Hamming distances to measure the dissimilarity between patterns and their centres. The WKMIX is improved over the KMIX algorithm, and utilises attribute weights derived from mutual information values calculated based on a combination of Baye’s theorem, the entropy, and Kullback Leibler divergence. The research in this thesis suggests that a decision support system, for cardiovascular medicine, can be built utilising the studied risk prediction models and pattern recognition techniques. The same may be true for other medical domains.EThOS - Electronic Theses Online ServiceGBUnited Kingdo