Research & AI Analysis

Machine Learning Model Performance

The Random Forest (RF) classifier consistently delivered the highest accuracy across different feature sets:

• All 200 Variables: An RF model achieved 87.3% accuracy, demonstrating the potential of a comprehensive dataset.
• Top 26 Weighted Importance Variables: Accuracy significantly increased to 92.1%, indicating effective feature selection in enhancing model performance and reducing noise.
• Final 3 Key Variables: Using only the three most predictive biomarkers, the RF model maintained a robust 84.9% accuracy, highlighting the efficiency and clinical applicability of this minimal feature set for practical screening.

The stepwise binary logistic regression analysis confirmed these three features as highly significant, with a total explanatory power of 71.8% (Nagelkerke R²) for early PD classification.

Data-Driven Methodology for Robust Detection

The study employed a rigorous biomechanical analysis and machine learning pipeline:

• Data Acquisition: 106 participants (63 early-stage PD, 43 healthy controls) performed a self-selected speed sit-to-walk (STW) task. Three-dimensional motion capture, force plates, and surface electromyography (sEMG) collected comprehensive kinematic, kinetic, and neuromuscular data.
• Feature Extraction: A total of 200 variables were extracted across three STW task phases, including spatio-temporal, kinematic (e.g., joint ROM, COM speed), kinetic (e.g., COP-COM displacement, ground reaction forces), and neuromuscular (e.g., sEMG amplitude, asymmetry index, co-contraction index) parameters.
• Feature Selection: Weighted feature importance analysis using Random Forest (RF) and Extreme Gradient Boosting (XGBoost) identified the top 26 variables. Subsequent stepwise binary logistic regression further refined this to the final three most predictive biomarkers.
• Classification Model: Seven machine learning classifiers (Logistic Regression, K-Nearest Neighbors, Naïve Bayes, Linear Discriminant Analysis, Quadratic Discriminant Analysis, Support Vector Machine, and Random Forest) were evaluated using five-fold cross-validation, with Random Forest showing the best performance.

Enterprise AI in Clinical Practice

This research demonstrates the potential of AI and biomechanical analysis for early PD screening, offering a non-invasive and efficient method to assist clinical diagnosis:

• Early Detection Advantage: Identifying PD before overt motor symptoms is crucial. The STW-based biomarkers provide objective measures that can detect subtle motor control impairments missed by standard clinical assessments, facilitating earlier therapeutic interventions.
• Fall Risk Prediction: The identified biomarkers, particularly those related to COM speed, COP-COM displacement, and trunk ROM, are directly associated with postural stability and gait initiation challenges. These early biomechanical deteriorations indicate increased fall risk in early-stage PD, allowing for proactive fall prevention strategies.
• Streamlined Screening: Utilizing a minimal set of three features for robust classification makes this approach highly practical for clinical settings, reducing the complexity and cost associated with comprehensive assessments. It can serve as an objective tool for screening and monitoring disease progression.
• Personalized Treatment & Monitoring: The specific biomechanical insights can inform personalized physical therapy and rehabilitation programs tailored to address the unique motor control deficits observed in early PD patients, potentially improving treatment efficacy and quality of life.