The present study sought to identify relationships between musical features of music therapy improvisations and clients' level of mental retardation, using a computationally-based method of analysis. 216 improvisations, contributed by 50 clients, were collected in MIDI format. Clients were divided into four groups according to their level of diagnosed mental retardation: 1 = none, 2 = mild, 3 = moderate, 4 = severe or profound. 43 client-related musical features were automatically extracted from their improvisations in the MATLAB computer environment and entered into a series of linear regression analyses as predictors of clients'level of mental retardation. The final model, which contained nine significant musical variables, accounted for 67% of the variation in clients' level of mental retardation. Specifically, level of mental retardation was best predicted by temporal elements of the music relating to note duration, note density, articulation, and amount of silence.