An approach to modeling complex real-world data such as biomedical signals is to develop pattern recognition techniques and robust features that capture the relevant information. In this paper, we use a deep belief network (DBN) to predict subcortical structures of patients with Parkinson's disease based on microelectrode records (MER) obtained during deep brain stimulation (DBS). We report on experiments using a data set involving 52 MER for the structures: zona incerta (Zi), subthalamic nucleus (STN), thalamus nucleus (TAL), and substantia nigra (SNR). The results show that our chosen features and network architecture produces a 99.5% accuracy of detection and classification of the subcortical structures under study. Based on the results we conclude that deep belief networks could be used to predict subcortical structure–mainly the STN for neurostimulation.