| Abstract |
This study extensively investigates the use of Artificial Neural Networks (ANN) in the diagnostic process of clinical psychiatry and focuses on developing improved methods based on designing efficient ANN systems. Psychiatric diagnosis is considered as the classification process of structured clinical interviews following specific schedules, although alternative considerations based on the evaluation of physical measurement are also possible. The 140-item clinical interviews used here are based on the Present State Examination (PSE) schedule. It is attempted to analyse the specific characteristics of this problem and design ANN of improved performance to solve it. The problem is considered in two different ways, concerning class labelling, according to two of the most widely accepted methodologies for defining mental disorders (ICD, PSE/CATEGO) and its specific characteristics are analysed. It is attempted to classify five main categories of psychiatric diseases as well as the non-abnormality class. The proposed ANN systems of the Multilayer Perceptron (MLP) type have specified architectures defined by incorporating clinicians' knowledge and they are compared with conventional MLPs. They are trained with the standard on-line Back propagation algorithm. In the sequel, superior solutions are sought by effectively reducing the input space dimensionality, based on clinicians' knowledge and by designing novel efficient task specific ANN architectures corresponding to this problem reformation. These last systems are compared with several other ANN techniques, including MLPs trained with Weigend's weight elimination approach, as well as with several traditional statistical pattern recognition methods, like multivariate discriminant analysis, k-NN and Euclidean distance classifier. The comparison criteria include percentage agreement, kappa and Y, which are the most popular in the clinical psychiatry literature. All the methods employed in this study are carefully and extensively compared within a consistent statistical framework using the cross-validation methodology. It is illustrated that constrained ANN architectures, when the largest possible amount of expert knowledge is incorporated in their design, can yield improved and stable generalization capability results, the best among all the methods involved in terms of the trade-off between maximum classification performance average and variance, as well as in terms of training reliability, related to the histogram of classification performance during training. Therefore, they represent a framework suitable for psychiatric modelling, compared to conventional statistical pattern recognition techniques, since their improved performance indicates that a mapping could be established between psychiatric diagnosis processing rules and ANN architectures. The unsupervised classification methods used are mainly Kohonen's SOM as well as k-means and several statistical hierarchical clustering techniques. It is found that the clustering performance of SOM is superior to the one of all the other techniques which actually fail to identify homogeneous groups of cases. This study clearly shows the feasibility of successfully employing artificial neural networks in interview based psychiatric diagnosis, despite the very large dimensionality of the pattern space, providing this way a strong experimental evidence that they do not significantly suffer from the curse of dimensionality. It is concluded that ANN of the MLP type, particularly the constrained ones, clearly outperform by far all the other techniques, especially the ones based on local classification methodologies, in terms of agreement with desired diagnoses. Several hypotheses are suggested to explain this performance, which are evaluated by direct application to an artificial problem with similar characteristics.
|
Collections
