A Double-Hurdle Quantification Model for Freezing of Gait of Parkinson’s Patients

Neurology Information Science Artificial Intelligence Medicine
Parkinson's disease Freezing of gait Zero-inflated data distribution Double-hurdle model Random forest algorithm Mixed-task learning mrmr

Research on Quantitative Model for Freezing of Gait in Parkinson’s Patients

Background Introduction

Parkinson’s Disease (PD) is a common neurodegenerative disease, accompanied by complex motor disorders. In the later stages of Parkinson’s disease, the phenomenon of “Freezing of Gait” (FOG) becomes particularly prominent. FOG refers to a transient phenomenon where patients suddenly cannot start or continue walking during the walking process. This phenomenon not only increases the risk of falls but also significantly reduces the patient’s mobility, severely affecting their quality of life. Therefore, accurately quantifying the severity of FOG is crucial for helping clinicians manage this symptom and mitigate its impact. Research Process

Currently, the commonly used New Freezing of Gait Questionnaire (NFOG-Q) in clinical practice mainly relies on patient self-reporting and physician experience for assessment. However, this assessment method is subjective and uncertain, unable to provide precise and detailed quantitative results. With the advancement of technology, research based on instrumented gait analysis has gained attention, but most studies often overlook fine-grained assessments when quantifying the severity of FOG.

Source of the Paper

This paper was written by several scholars, including Xu Ningcun, Wang Chen, Peng Liang, Zhou Xiaohu, Chen Jingyao, Zhi Cheng, and Hou Zengguang. It was published in the 2023 issue of the journal “IEEE Transactions on Biomedical Engineering.” The authors are from the Faculty of Innovative Engineering, Macau University of Science and Technology, and the Institute of Intelligent Automation.

Research Methods and Process

This paper proposes a double-hurdle model to quantify the severity of FOG in Parkinson’s patients using typical spatiotemporal gait features. Additionally, a new multi-output random forest algorithm (MGWRF) is introduced to further enhance model performance. The study includes six experiments using a public Parkinson’s disease gait database for validation. The research process is as follows:

a) Research Process

  1. Gait Feature Data Collection and Preprocessing

    • Gait data were collected using force platforms and multiple camera devices. The database includes 13 PD patients with FOG symptoms and 13 PD patients without FOG symptoms, recording their gait features in “medication ON” and “medication OFF” states.

  2. Feature Extraction and Standardization

    • The initial gait features include 19 indicators such as standing time, swing time, step length, and walking speed. The original gait database was expanded, normalized, and dimensionally reduced to obtain standardized gait feature data.

  3. Gait Feature Selection

    • The Maximum Relevance Minimum Redundancy (MRMR) algorithm was used to select the most important gait features from the standardized data to construct the input feature set.

  4. Construction of the Double-Hurdle Model

    • The first hurdle of the model is used to identify FOG patients, and the second hurdle is used to score the severity of FOG in patients.

  5. Application of the MGWRF Algorithm

    • Bias differential information entropy quantification technology was introduced to improve the decision tree node splitting standard and leaf node prediction mechanism for multitask processing, including patient classification and FOG scoring.

  6. Analysis of the Influence of Medication State on Gait Patterns

    • Key indicators of medication state in gait features were selected to analyze their impact on gait features under FOG conditions.

b) Major Research Results

The research results show that, within the hyperparameter-independent framework of the double-hurdle model, the MGWRF algorithm achieved the highest correlation coefficient of 0.972 and the lowest root mean square error of 2.488. Moreover, the medication state significantly affected the gait patterns of patients:

  1. In the “medication OFF” state, FOG symptoms were significantly enhanced.
  2. Comparing the “medication ON” and “medication OFF” states, step length and walking speed significantly decreased in FOG patients, while gait variability (such as variability in gait timing and length) significantly increased.

c) Conclusion and Research Value

  1. Scientific Value

    • This study is the first to use a double-hurdle model for fine-grained quantification of FOG severity in Parkinson’s patients, providing a more accurate assessment tool.

  2. Application Value

    • The research results help clinicians more accurately manage and treat FOG symptoms, improving the quality of life for Parkinson’s patients.

  3. Innovative Highlights

    • The proposed new MGWRF algorithm, combined with multitask learning, significantly enhances the processing and analysis performance of gait data.

  4. Novelty of Methods

    • Using bias differential information entropy to quantify node information gain, combined with task-sharing weight mechanisms, enhances the model’s performance in handling high-dimensional data.

d) Research Highlights

  1. Addressing the Zero-Inflated Data Distribution Problem

    • The zero-inflated data distribution problem in FOG severity scoring is effectively addressed by the double-hurdle model.

  2. Innovation in Modeling Methods

    • The MGWRF algorithm utilizes bias differential information entropy quantification, enhancing the performance of the model in multitask processing.

e) Other Valuable Information

Further research validated the impact of medication state on gait features, indicating that FOG symptoms are more easily observed in the “medication OFF” state. This finding suggests that clinical studies should consider the impact of medication state on gait analysis to ensure the accuracy of assessments.

Conclusion

The double-hurdle model proposed in this paper, through fine-grained quantification and multitask learning, overcomes the shortcomings of traditional methods and improves the accuracy of FOG assessment. The research results provide new perspectives and methods for managing and treating Parkinson’s disease and have significant reference value for clinical research and applications related to FOG symptoms. Future research can further validate and expand the model’s applicability by collecting more patient data and incorporating multimodal data.