Predicting music therapy clients' type of mental disorder using computational feature extraction and statistical modelling techniques

Abstract

Background. Previous work has shown that improvisations produced by clients during clinical music therapy sessions are amenable to computational analysis. For example, it has been shown that the perception of emotion in such improvisations is related to certain musical features, such as note density, tonal clarity, and note velocity. Other work has identified relationships between an individual's level of mental retardation and features such as amount of silence, integration of tempo with the therapist, and amount of dissonance. The present study further develops this work by attempting to predict music therapy clients' type of mental disorder, as clinically diagnosed, from their improvisatory material. Aim. To predict type of mental disorder from computationally-extracted musical features of music therapy improvisations. Method. Two hundred and sixteen music therapy improvisations, obtained from seven music therapists' regular sessions with their clients, were collected in MIDI format. A total of fifty clients contributed musical material, and these clients were divided into three groups according to their clinical diagnosis: RET (mentally retarded), DEV (developmental disorder), and NEU (neurological disorder). The improvisations were subjected to a musical feature extraction procedure in which 43 musical features were automatically extracted in the MATLAB computing environment. These features were then entered into a discriminant function analysis as predictors of type of mental disorder. Results. The analysis produced two significant discriminant functions, and the emergent model correctly classified 80% of clients. Significant predictor variables fell into three main categories: those relating to pitch, those relating to temporal aspects, and those relating to tonal clarity and dissonance. Conclusions. The present study suggests that an individual's type of mental disorder can be predicted from a statistical analysis based upon the computational extraction of detailed musical features from their improvisatory material. As such, it offers further evidence for the usefulness of computational music analysis in music therapy. © 2009 Springer-Verlag.