Explaining deep learning models for age-related gait classification based on acceleration time series

Area of Science:

• Biomedical Engineering
• Data Science
• Gerontology

Background:

• Gait analysis is crucial for monitoring older adult health.
• Advancements in sensor technology generate big data for gait analysis.
• Deep learning (DL) shows promise but lacks transparency for clinical use.

Purpose of the Study:

• Enhance transparency in DL-based gait classification for aged-related patterns.
• Utilize Explainable Artificial Intelligence (SHAP) to interpret DL models.
• Improve clinical applicability of AI in gait analysis.

Main Methods:

• Cross-sectional study with 244 participants (adults and older adults).
• Used accelerometers on L3 during a 3-min walk.
• Trained Convolutional Neural Network (CNN) on 1-stride and Gated Recurrent Unit (GRU) on 8-stride data.
• Applied SHAP for model explanation.

Main Results:

• CNN achieved 81.4% accuracy (AUC 0.89); GRU achieved 84.5% accuracy (AUC 0.94).
• SHAP identified vertical and walking direction data around heel contact as most important.
• GRU's analysis considered inter-stride variations, unlike CNN.

Conclusions:

• CNN classifies gait based on single-stride data; GRU uses inter-stride relationships.
• Heel contact data is critical for differentiating adult and older adult gait patterns.
• Explainable AI (SHAP) provides insights into DL models for gait analysis.