Autoencoder-Enhanced Convolutional Neural Networks for Plantar Pressure-Based Gait Pattern Recognition: Model...

Area of Science:

• Biomechanics
• Machine Learning
• Medical Imaging

Background:

• Plantar pressure imaging offers insights into gait biomechanics for assessment and recognition.
• High dimensionality and nonlinearity of plantar pressure data challenge traditional machine learning methods.
• Developing robust gait recognition models requires advanced feature extraction techniques.

Purpose of the Study:

• To create a gait pattern recognition model using an autoencoder (AE)-enhanced convolutional neural network (CNN) with plantar pressure data.
• To assess the performance of the AE-CNN model against other deep learning and classical machine learning approaches.
• To enhance the accuracy and reliability of gait analysis through advanced computational methods.

Main Methods:

• Collected plantar pressure data from 13 healthy volunteers during treadmill walking using an in-shoe system.
• Converted pressure data into frame-wise plantar pressure images for analysis.
• Compared an encoder-augmented CNN model against a lightweight CNN and an AE-CNN cascade model, evaluating performance using accuracy, precision, recall, and F1-score.

Main Results:

• The encoder-augmented CNN model achieved the highest performance with an F1-score of 96.20%.
• This model outperformed the lightweight CNN (94.44% F1-score) and AE-CNN cascade (92.45% F1-score).
• Stable training behavior and consistent classification performance were confirmed through confusion matrices and learning curves.

Conclusions:

• Combining representation learning (AE-based compression) with CNN classification enhances gait pattern recognition from plantar pressure images.
• This pilot study in healthy participants suggests a promising approach for gait analysis.
• Future research should focus on generalizability in diverse cohorts and model interpretability.