Parkinson's disease is a chronic neurodegenerative disorder that severely impairs patients' quality of life. Currently, the waiting lists for a neurologist are quite long, and during this period, patients are without adequate therapy to alleviate their symptoms. Therefore, we have developed an automatic method to determine the level of motor impairment or bradykinesia of Parkinson's disease based on a tapping test, thus enabling faster diagnosis. We collected 183 tapping videos recorded with a smartphone in everyday environments, which were assessed by a neurologist into 5 classes of the MDS-UPDRS scale. For hand detection we used MediaPipe Hand, which returns a time series of the hand skeleton. For classification, we took two different approaches. First, we constructed features from the hand skeleton time series, once strictly following the MDS-UPDRS scale, and another time not strictly adhering to it. These features were then used in classifiers and achieved 61 \% accuracy and 0,62 F1 score using a multi-layer perceptron. In the second approach, we used the time series of thumb-pointer distances directly in a fully convolutional neural network achieving 77 \% accuracy and 0,75 F1 score. We also created a tool for visualizing tapping and displaying key data.
|
