Selective Motor Entropy Modulation and Targeted Augmentation for the Identification of Parkinsonian Gait Patterns...

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Sensor Technology

Background:

• Parkinsonian gait exhibits impaired motor adaptability and reduced variability.
• Current data augmentation methods may distort pathological gait dynamics.
• Need for physiologically informed augmentation strategies in gait analysis.

Purpose of the Study:

• To investigate if preserving healthy motor variability while augmenting pathological gait signals enhances classification models.
• To evaluate the impact of selective augmentation on gait pattern recognition.
• To improve the robustness and physiological coherence of gait classification.

Main Methods:

• Collected kinematic data from Parkinsonian patients and healthy individuals using wearable inertial sensors on a specialized platform.
• Applied a selective data augmentation technique (smooth time-warping) exclusively to pathological gait segments.
• Utilized a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architecture for model evaluation.

Main Results:

• Selective augmentation of pathological gait signals yielded the highest classification performance (94.1% accuracy, AUC=0.97).
• Optimal augmentation was constrained by physiologically plausible temporal dynamics, with performance decreasing beyond this range.
• Achieved balanced sensitivity (93.8%) and specificity (94.3%).

Conclusions:

• Physiology-informed, selective data augmentation improves gait pattern classification accuracy, especially with limited data.
• Respecting intrinsic motor variability differences is crucial for effective augmentation strategies in clinical gait analysis.
• Further research with diverse cohorts and subject-independent validation is needed for clinical generalizability.