Multi-scale and Multi-level Feature Assessment Framework for Classification of Parkinson’s Disease State from Short-term Motor Tasks

Artificial Intelligence Neurology Information Science Life Sciences Bioengineering Medicine
Parkinson’s disease feature assessment machine learning wearable devices short-term motor tasks

Academic Background

Parkinson’s Disease (PD) is the second most common chronic neurodegenerative disease, primarily affecting individuals aged 65 and above. With the global population aging, the prevalence of Parkinson’s disease is projected to increase from 7 million in 2015 to 13 million by 2040. Currently, the diagnosis of Parkinson’s disease mainly relies on clinical questionnaires and motor diaries, which are time-consuming and subject to significant subjective bias. In recent years, with the integration of wearable technology and machine learning methods, researchers have begun exploring the quantification of motor symptoms to assist in the diagnosis of Parkinson’s disease. However, the effectiveness of these technologies is influenced by environmental settings, making it difficult to apply them widely in real-world scenarios. Therefore, this study aims to propose an effective feature assessment framework to automatically evaluate the severity of motor symptoms in Parkinson’s disease through short-term motor tasks and classify them in real-world settings.

Source of the Paper

This research was conducted by a team from the University of Sheffield and Yunnan University, with primary authors including Xiyang Peng, Yuting Zhao, Ziheng Li, Xulong Wang, Fengtao Nan, Zhong Zhao, Yun Yang, and Po Yang. The paper was published on June 21, 2024, in IEEE Transactions on Biomedical Engineering, titled “Multi-Scale and Multi-Level Feature Assessment Framework for Classification of Parkinson’s Disease State from Short-Term Motor Tasks.”

Research Process and Details

Research Process

  1. Data Collection and Preprocessing
    The study was conducted at the First People’s Hospital of Yunnan Province, involving 100 Parkinson’s disease patients and 60 healthy controls (25 young individuals and 35 elderly individuals). Participants wore Shimmer wearable devices (sampling rate 200Hz) and performed 14 short-term motor tasks, each lasting 20 to 50 seconds. The data underwent normalization, filtering, and segmentation for subsequent analysis.

  2. Feature Extraction
    The study proposed a multi-scale (sample-level and segment-level) and multi-level (time-domain, frequency-domain, spectral-domain, autocorrelation-domain) feature assessment framework. Specific features included peak numbers, sample entropy, autocorrelation coefficients, etc. After feature extraction, SHAP (Shapley Additive Explanations) values were used for feature selection, and various machine learning methods validated the effectiveness of the features.

  3. Feature Selection and Classification
    The study employed classifiers such as LightGBM, SVM, KNN, XGBoost, logistic regression, and convolutional neural networks (CNN) for classifying Parkinson’s disease states. Model performance was evaluated using Leave-One-Subject-Out (LOSO) cross-validation.

Main Results

  1. Motor Symptom Recognition
    The study successfully identified motor symptoms of Parkinson’s disease, such as tremors and bradykinesia, with tremor recognition sensitivity reaching 88%. Through analysis of right-hand rotation and sitting tasks, sample-level features (e.g., amplitude area, normal/abnormal peak numbers) demonstrated high accuracy in detecting motor fluctuations.

  2. Classification of Parkinson’s Disease Severity
    The study classified the severity of Parkinson’s disease across multiple short-term motor tasks, with the “walking around” (WA) task performing best, achieving an accuracy of 71.58%. Using SHAP value feature ranking, 31 key features were ultimately selected for fine-grained classification.

  3. Feature Assessment and Classifier Performance
    The study found that sample-level features, time-domain features, and autocorrelation-domain features performed well in early Parkinson’s disease detection, while spectral-domain features were more effective in fine-grained classification. The LightGBM classifier excelled in multiple tasks, particularly in fine-grained classification.

Conclusions and Significance

This study proposes, for the first time, a multi-scale and multi-level feature assessment framework for automatically quantifying motor symptoms and their severity in Parkinson’s disease through short-term motor tasks. The framework demonstrated high effectiveness in real-world data, offering new possibilities for self-assessment in Parkinson’s disease. The results indicate that by analyzing key features from short-term motor tasks, motor symptoms of Parkinson’s disease can be effectively identified and classified by severity. This framework not only holds significant scientific value but also provides essential tools for clinical diagnosis and home-based self-monitoring of Parkinson’s disease.

Research Highlights

  1. Multi-Scale and Multi-Level Feature Assessment: The study, for the first time, extracted and analyzed features at multiple levels, including sample-level and segment-level, time-domain, frequency-domain, spectral-domain, and autocorrelation-domain, providing a comprehensive feature assessment framework for automatic classification of Parkinson’s disease.
  2. Validation with Real-World Data: The study collected data in a real clinical environment, validating the effectiveness of the framework and providing strong support for practical applications in Parkinson’s disease.
  3. Efficient Feature Selection and Classification: Through SHAP value feature ranking and the LightGBM classifier, the study successfully screened key features and achieved high accuracy in fine-grained classification tasks.

Other Valuable Information

The study also found that different motor tasks contribute differently to the classification of Parkinson’s disease states. For example, the “walking around” task performed best in fine-grained classification, while the “drinking” task excelled in early detection. These findings provide direction for future research, allowing for the selection of appropriate motor tasks for Parkinson’s disease assessment based on specific needs.

Through this study, the researchers not only proposed a new feature assessment framework but also provided important theoretical support and practical guidance for the automatic diagnosis and severity classification of Parkinson’s disease. In the future, with the accumulation of more data and optimization of algorithms, this framework is expected to find wide application in clinical and home environments, offering more convenient and accurate diagnostic services for Parkinson’s disease patients.