We propose adaptive learning methods for identifying different behavioral tasks of patients with Parkinson's disease (PD). The methods use local field potential (LFP) signals that were collected during Deep Brain Stimulation (DBS) implantation surgeries. Using time-frequency signal processing methods, features are first extracted and then clustered in the feature space using two different methods. The first method requires training and uses a hybrid model that combines support vector machines and hidden Markov models. The second method does not require any a priori information and uses Dirichlet process Gaussian mixture models. Using the DBS acquired signals, we demonstrate the performance of both methods in clustering different behavioral tasks of PD patients and discuss the advantages of each method under different conditions.