Evaluation of objective features for classification of clinical depression in speech by genetic programming

Abstract

This paper presents the results of applying a Genetic Programming (GP) based feature selection algorithm to find a small set of highly discriminating features for the detection of clinical depression from a patient's speech. While the performance of the GP-based classifiers was not as good as hoped for, several Bayesian classifiers were trained using the features found via GP and it was determined that these features do hold good discriminating power. The similarity of the feature sets found using GP for different observational groupings suggests that these features are likely to generalize well and thus provide good results with other clinical depression speech databases. © 2007 Springer-Verlag Berlin Heidelberg.