JMIR Publications

ABSTRACT

Background:

Foot pressure imaging is a non-invasive and stable signal for gait analysis and pathology screening, but manual features and classical machine learning struggle with high-dimensional, nonlinear data.

Objective:

This study aims to develop and evaluate a deep learning–based foot pressure recognition system by integrating convolutional neural network (CNN) architectures with autoencoder (AE) feature compression, and to compare its performance with traditional classifiers.

Methods:

Thirteen healthy adults performed three gait conditions (slow, brisk, uphill). Tekscan F-scan data were converted to 64×64 grayscale images with normalization and light augmentation. Three models were built—Light CNN, AE-CNN Cascade, and Encoder-augmented CNN—and benchmarked against KNN, SVM, and Random Forest using 9-fold cross-validation. Metrics included accuracy, precision, recall, and F1-score.

Results:

A total of 6,994 foot pressure frames were analyzed. The Encoder-augmented CNN achieved the best performance across all metrics, with an F1-score of 96.21% and stable convergence under batch normalization optimization. Among traditional methods, the SVM with RBF kernel achieved the highest accuracy. Comparative results confirm that AE-based feature compression significantly enhances CNN classification, outperforming both standard CNNs and classical classifiers.

Conclusions:

Integrating AE-based dimensionality reduction with CNNs enables accurate, robust foot-pressure recognition, supporting applications in real-time gait monitoring and pathological gait screening. Clinical Trial: Not applicable.