Parkinson’s disease is one of the most commonly studied neurological disorders in
the field of machine learning. Timely patient care and assessment of the severity
level significantly contribute to appropriate therapy, which is often hindered by
long waiting times. We developed a classification model based on the criteria
described in the Movement Disorder Society-Unified Parkinson’s Disease Rating
Scale (MDS-UPDRS). According to the descriptions in the MDS-UPDRS scale,
we generate synthetic finger-tapping signals, which we use to train deep machine
learning models; the developed models are then applied to real data obtained
from video recordings. The scores define three key anomalies: decrease in amplitude,
decrease in speed, and signal interruptions, which we mathematically
described and separately modeled using three simple autoencoders. The latent
spaces of the autoencoders served as input data for the k-nearest neighbors (kNN)
method, which we used to determine the severity level based on each anomaly.
For new cases, we do not perform additional training; instead, we compare the
observed finger-tapping signal with a set of synthetically generated signals and
use the kNN method to find which predefined signal the observed one is closest
to, enabling assessment consistent with the MDS-UPDRS criteria. The highest
accuracy was achieved by the model predicting the decrease in speed, although
the numerical definition of this anomaly is challenging due to the textual description
in the MDS-UPDRS scale. A total of 183 finger-tapping videos were
collected, from which tapping signals were extracted using the MediaPipe Hand
tool. These signals enabled the generation of a synthetic dataset and the testing
of the classification approach. The autoencoders and kNN methods contribute
to understanding the relationship between mathematically defined anomalies in
finger movement signals and the MDS-UPDRS scores.
|
