It is well known that long-term use of levodopa by patients with Parkinson's disease causes dyskinesia. Several methods have been proposed for the automatic, unsupervised detection and classification of levodopa induced dyskinesia. Recently, we have demonstrated that neural networks are highly successful to detect dyskinesia and to distinguish dyskinesia from voluntary movements. The aim of this study was to use the trained neural networks to extract parameters, which are important to distinguish between dyskinesia and voluntary movements. Thirteen patients were continuously monitored in a home-like situation performing in about 35 daily life tasks for a period of approximately 2.5 h. Behavior of the patients was measured using triaxial accelerometers, which were placed at six different positions of the body. A neural network was trained to assess the severity of dyskinesia. The neural network was able to assess the severity of dyskinesia and could distinguish dyskinesia from voluntary movements in daily life. For the trunk and the leg, the important parameters appeared to be the percentage of time that the trunk or leg was moving and the standard deviation of the segment velocity of the less dyskinetic leg. For the arm, the combination of the percentage of time, that the wrist was moving, and the percentage of time, that a patient was sitting, explained the largest part of the variance of the output. Dyskinesia differs from voluntary movements in the fact that dyskinetic movements tend to have lower frequencies than voluntary movements and in the fact that movements of different body segments are not well coordinated in dyskinesia.