Parkinson's disease is the second most common degenerative disease. It is incurable, however with early discovery, its progression can be slowed. Early signs of the disease are often noticed by family physicians, but due to long waiting times for neurologists, diagnosis and consequently treatment can be delayed. Automating the prediction of disease progression could reduce those waiting times, which is the motivation behind this thesis. One of the methods to determine the level of disease progression is by using a finger tapping test, which is part of the MDS-UPDRS grading scale. In this work, we used video recordings in which patients performed the finger tapping test. These recordings were evaluated by a physician on a scale from 0 to 4 according to the criteria defined by the scale. Hand detection was performed using MMPose and features were extracted from the obtained keypoints. For classification we used Gaussian Processes, where we compared the use of different covariance functions (kernels). The best result was achieved using the Matern 3/2 kernel, with the model reaching 66% accuracy and an F1 score of 0,66.
|
