Tremor detection using motion filtering and SVM

The hand tremor is one of the most common motion disorders caused by various neurological diseases. Currently diagnostic procedures for tremor evaluation are subjective, and there are no examinations available that can accurately indicate whether tremors are present in a patient's daily life. Early detection of tremor is extremely important for the cure of the disease that causes the tremor. Thus, in this study we aim to develop a computational method based on machine learning that can automatically detect hand tremors from a video of a patient when the patient is doing his/her daily activities. The main challenge in tremor detection is that motion is very subtle, and the signature of the motion can vary from patient to patient. We first generate a training data set consisting of 173 simulated tremor/non-tremor video files. They contain very subtle and less discriminative motions. The main contributions of our study are the personalized skin detection, motion filtering and feature extraction pipeline. We evaluated our method through leave-one-out cross validation (LOOCV) testing and showed that in preliminary tests our method achieved 95.4% recognition accuracy.