Neural networks in multidimensional problems: a case study for questionnaire classification

This paper focuses on the classification problem of high dimensional patterns and especially of psychiatric questionnaires. The use of artificial neural networks (ANN) in interview based psychiatric diagnosis is investigated and innovative constrained architectures are proposed. The network inputs come from a large database containing 140-item structured clinical interviews based on the present state examination schedule. It is attempted to classify five main psychiatric diseases and the non-disease case. Constrained task-specific architectures, as well as non-constrained topologies trained with the online backpropagation (BP) algorithm and other learning rules, are utilised. The overall average classification accuracy achieved by the best constrained architecture is 85.92%. This study clearly shows the feasibility of successfully employing efficient ANN models in psychiatric diagnosis. Finally, some interesting hints about the nature of psychiatric diagnosis classification problem are provided using ANNs.