Design of parametric fusion classifiers

Classifier fusion and data fusion methods are formulated to improve the performance of pattern classifiers. Three distinct classification problems are considered for the development of the classifiers. The first problem involves the classification of multichannel evoked potentials (EPs). A fusion rule is formulated to classify single-trial and averaged EPs using the classification results from all the EP channels. Problems with estimating parameters of EPs averaged from a finite set of single-trial EPs are identified and it is systematically shown that the parameters can be estimated directly from the single-trial ensemble. It is shown that through single-trial amplitude and slope normalization, the intra-class variability is decreased and the inter-class separation is increased. Problems relating to the partitioning of averaged EPs into design and test sets are identified and a random sampling approach is developed to robustly design and evaluate the performance of EP classifiers. Features with high inter-class separations are selected through the discrete Karhunen-Loeve transform (DKLT). A decision theory approach is used to develop classifiers. Experiments show that the performance of the classification fusion methodology developed is consistently superior when compared with the performance of selecting the best channel.; Methods are formulated to fuse the risk factors of multifactorial diseases into a single feature vector. The need for normalization is also identified. The need for transforming the normalized feature vector to facilitate the development of parametric classifiers is also identified. The first data fusion method is formulated specifically for 2-class classification problems aimed at detecting the presence or absence of a multifactorial disease. The method developed is applied to predict the occurrence of gout. Because the method developed for the 2-class case cannot be directly extended to several classes, the next data fusion method is developed for multi-class classification problems. A pairwise rank ordering and rank-sum rule is formulated to select features with high inter-class separations between the multiple classes. The method developed is applied to predict Alzheimer's disease (AD) from a set of 23 risk factors. The prediction problem is formulated as a 3-class classification problem in which the 3 classes are AD-likely, AD-possible, and AD not-likely.