Patients with Parkinson's disease often experience significant changes in the severity of dyskinesia when they undergo titration of their medications. Dyskinesia is marked by involuntary jerking movements that occur randomly in a burst-like fashion. The burst-like nature of such movements makes it difficult to estimate the clinical scores of severity of dyskinesia using wearable sensors. Clinical observations are generally made over intervals of 15-30 s. On the other hand, techniques designed to estimate the severity of dyskinesia based on the analysis of wearable sensor data typically use data segments of approximately 5 s. Consequently, some data segments might include dyskinetic movements, whereas others might not. Herein, we propose a novel method suitable to automatically select data segments from the training dataset that are marked by dyskinetic movements. The proposed method also aggregates results derived from the testing dataset using a machine learning algorithm to estimate the severity of dyskinesia from wearable sensor data. Results obtained from the analysis of sensor data collected from seven subjects with Parkinson's disease showed a marked improvement in the accuracy of the estimation of clinical scores of dyskinesia.