Autoencoder-Enhanced Convolutional Neural Networks for Plantar Pressure-Based Gait Pattern Recognition: Model Development and Cross-Validated Evaluation Study.

BACKGROUND

Plantar pressure imaging is a stable modality that reflects gait-related biomechanical characteristics and has been used increasingly for gait assessment and recognition. However, plantar pressure images are high dimensional and nonlinear, making manual feature engineering and conventional machine learning insufficient to capture discriminative patterns.

OBJECTIVE

This study aimed to develop a gait pattern recognition model based on plantar pressure using an autoencoder (AE)-enhanced convolutional neural network (CNN) and to evaluate its performance against baseline deep learning and classical machine learning approaches.

METHODS

A total of 13 healthy volunteers (aged 18-24 years) were recruited. Plantar pressure data were collected during treadmill walking using an in-shoe pressure measurement system and converted into frame-wise plantar pressure images. We compared a lightweight CNN (Light CNN), an AE-CNN cascade model, and an encoder-augmented CNN with an additional bottleneck layer. Model development used participant-wise data partitioning, and performance was evaluated using accuracy, precision, recall, and F1-score.

RESULTS

The proposed encoder-augmented CNN achieved the best overall performance (F1-score=96.20%), outperforming the Light CNN (F1-score=94.44%) and AE-CNN cascade (F1-score=92.45%). Confusion matrices and learning curves further indicated stable training behavior and consistent classification performance across gait patterns.

CONCLUSIONS

Integrating representation learning (AE-based compression) with CNN-based classification improved the recognition of gait patterns from plantar pressure images. This pilot study included only healthy participants. Future work should validate generalizability in larger and clinically diverse cohorts and further investigate participant-level evaluation and model interpretability, as well as deployment feasibility.