The results of a study conducted on the subjects with Parkinson's disease implanted with deep-brain stimulators, by collecting accelerometer data from patient-worn sensors, are presented. The accelerometer time series is segmented and different combinations of accelerometer data are used to predict the severity of different disease symptoms. The data-feature structure is visually examined using Sammons mapping that measured change in feature values from pretest trail to trails performed with simulator turned off. The results show that average prediction errors ranged from approximately 5% to 23% for one-second epochs. The cluster distance measure shows a significant increase between the pretest cluster and the first trail with the stimulator off. A change in the bradykinesia clinical score is observed from 0 to 3 when the stimulator is off, while the cluster distance measure decrease indicating that the accelerometer features are moving closer to those of the pretest conditions.