Explainable AI for gait speed analysis from multimodal data fusion.

Area of Science:

• Biomechanics
• Machine Learning
• Data Science

Background:

• Gait speed analysis is crucial for healthcare, rehabilitation, and human-robot interaction.
• Existing methods face challenges in accurately capturing complex gait dynamics.

Purpose of the Study:

• To develop an advanced framework for gait speed classification using multimodal data fusion and deep learning.
• To optimize the framework using Layer-wise Relevance Propagation (LRP) for enhanced feature selection and model robustness.

Main Methods:

• Integrated full-body motion capture, electromyography (EMG), and force plate data from 50 adults across 4 datasets.
• Proposed and benchmarked novel deep learning architectures (CNN+LSTM, Multi-stream CNN) against TCNs, Transformers, GRUs, and statistical classifiers.
• Employed three distinct cross-validation strategies, including subject-based splitting, to ensure rigorous model evaluation.

Main Results:

• The Multi-stream Quads CNN model achieved the highest F1 scores across all experiments, reaching up to 98%.
• The framework demonstrated superior performance compared to traditional statistical classifiers and other deep learning models.
• Layer-wise Relevance Propagation (LRP) successfully identified critical features, validating the model's robustness and durability through perturbation analyses.

Conclusions:

• The proposed multimodal data fusion and deep learning framework offers a highly accurate and resilient solution for gait speed classification.
• LRP-driven feature optimization enhances model performance and interpretability.
• This advanced tool has significant potential for applications in healthcare, rehabilitation, and autonomous systems.