Assessment of Parkinson's Motor Severity by Multi-feature and Multi-scale Motion Convolutional Neural Network with Positional Encoding

Parkinson's disease is one of the most common neurodegenerative diseases. Its main clinical manifestations are a variety of movement disorders, including abnormal gait, which seriously affect the performance of motor function, leading to shaking hands at rest, bradykinesia, myotonia, and difficulty in maintaining postural balance. The current work focuses on the methods of wearable devices and non-wearable devices to evaluate sports injuries. However, the network models based on the Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS) to quantify the severity of Parkinson's disease injuries have problems such as incomplete generalization of input features, insufficient feature extraction of network models, and imbalance between recognition accuracy and model size. A lightweight and high-precision convolutional neural network model is essential for accurate and quantitative assessment of disease progression and for slowing or preventing further disease progression. Regarding the above issues, we use a computer vision-based model to observe video recordings of the human body, extract their 3D body skeletons, track them through time, and input them into our proposed a Triple-feature Triple-motion Convolutional Neural Network (PE-TTNET) with Positional encoding for training. We position encode the triple scale features and classify the gait severity of PD patients based on MDS-UPDRS. The experimental results indicate that PE-TTNET achieved better results compared to previous network models in testing the clinical dataset we collected, with higher Fl-score and average accuracy in classifying the severity level of Parkinson's disease.