The dynamic nature of Parkinson's disease motor disorders makes clinical assessments based on the UPDRS rating scale a time-consuming and expert-dependent process. Although the use of wearable sensors and machine learning algorithms provides an objective solution for analyzing these disorders, the complexity of motor patterns and the scarcity of labeled data have limited the efficiency of these models, particularly in the application of deep learning methods.

In this study, a two-stage framework is proposed to diagnose and determine the severity of Parkinson's disease motor disorders, relying on unsupervised feature extraction and similarity-based classification. In the first stage, a hierarchical feature extraction model based on the unsupervised HUF framework is utilized. By combining convolutional networks and autoencoders, this model systematically performs the extraction and fusion of motor features at three levels: sensor channels, co-located sensors, and the entire set of sensors involved in each movement. In the second stage, to overcome the challenge of limited labeled data, an approach based on meta-learning and Siamese networks is introduced for feature classification. Rather than using data labels directly, this method maps motor features into a novel representation space and makes decisions by measuring the similarity between the feature distributions and a predefined set of reference features, ultimately selecting the class with the highest similarity.

The proposed model was evaluated on the PD-BIOSTAMPRC21 dataset for three motor tasks, including walking, pronation-supination of right and left hand. For Parkinson's detection, the results achieved accuracies of 90%, 93%, and 86%, respectively. Additionally, for determining the disease severity level, by adopting a hierarchical one-vs-one classification strategy, accuracies of 79%, 74%, and 77% were obtained for the three mentioned movements, respectively. These results demonstrate that the proposed unsupervised framework has a high capability in extracting and classifying motor patterns using multi-channel data from wearable sensors.