Input representations and classification strategies for automated human gait analysis.

Area of Science:

• Biomechanics
• Machine Learning
• Clinical Data Analysis

Background:

• Quantitative gait analysis generates complex data challenging for clinicians.
• Machine learning offers potential for automated gait classification and data interpretation.
• Lack of consensus exists on optimal data preprocessing for accurate gait classification.

Purpose of the Study:

• To determine the optimal data aggregation and preprocessing workflow for enhancing gait classification accuracy.
• To compare different data fusion techniques and signal representations for machine learning models.

Main Methods:

• A sequential approach compared early and late data fusion methods.
• Investigated signal representations including relative changes and signal differences.
• Utilized a dataset of 910 subjects with gait disorders and healthy controls.
• Employed Principle Component Analysis (PCA), z-standardization, and Support Vector Machine (SVM) for classification.

Main Results:

• Late fusion, using majority voting on classifier predictions, yielded superior performance.
• Derived signal representations demonstrated an advantage in classification accuracy.
• The chosen machine learning pipeline effectively classified gait data.

Conclusions:

• Data preprocessing and aggregation methods significantly impact classification accuracy in gait analysis.
• Results provide a guideline for selecting appropriate techniques in future studies.
• Optimized methods enhance the clinical utility of automated gait classification.