The general workflow of this study is illustrated in Figure 1. The experimental design consisted of 3 major components. The first component involved the collection of plantar pressure response data required to generate plantar pressure images. The second component focused on the application of machine learning classifiers and a deep learning model based on a CNN. The third component addressed the AE, which was further investigated as a core part of this study. In the final stage of the experiments, the CNN and AE frameworks were integrated to perform classification and performance evaluation.

As shown in Figure 1, the experimental workflow highlights the sequential architecture of the study. In particular, the second component, which encompasses both classifiers and CNN-based models, is presented in separate blocks. The detailed structures of these blocks are illustrated in Figures 2 and 3.

In the classifier block diagram, the AE-CNN cascade is defined as an AE in which compressed data are modularized and sequentially integrated with a CNN. This design enables the model to be trained in a staged manner, thereby forming the AE-CNN cascade. Furthermore, because its data processing pathway resembles that of conventional classifiers, the AE-CNN cascade was grouped within the classifier block for consistency in experimental design.

In the deep learning block diagram, the encoder-augmented CNN is defined as a hybrid model in which the encoder component of the AE is directly integrated into the CNN classifier. This design enables end-to-end training, allowing the encoder and CNN to fuse into a unified framework. By adopting this hybrid strategy, the model is trained as an integrated architecture, which is referred to as encoder-augmented CNN.

A total of 13 healthy university student volunteers (aged 18-24 years) were recruited for this study. Exclusion criteria included current or previous foot ulcers, diabetes, vascular disease, hypertension, inability to walk independently for at least 10 minutes, and continued use of medications that could affect gait. Only participants who met all eligibility criteria were enrolled.

This study was approved by the Central Regional Research Ethics Committee of China Medical University, Taichung, Taiwan (approval CRREC-112-130). All participants received a full explanation of the study procedures and provided their written informed consent prior to participation. The data were deidentified for analysis and reporting. No financial compensation was provided to participants in this study.

Plantar pressure data were acquired using the Tekscan F-Scan in-shoe pressure measurement system (Figure 4 [18]). The F-Scan provides high-resolution plantar pressure distribution with real-time acquisition via a dedicated data cable, making it suitable for gait analysis, sports science, insole design, and clinical applications. The key functions adopted in this study are summarized in Table 1. In particular, the system’s real-time recording capability and high spatial resolution enabled the capture of subtle pressure changes, thereby improving the fidelity of the experimental dataset.

Plantar pressure responses were recorded using the Tekscan F-Scan in-shoe system while participants walked on a treadmill at a fixed speed. Each trial began with a 1-minute familiarization period to stabilize gait at the target speed. The F-Scan microsensors sampled plantar pressure at 25 Hz, and instantaneous pressures were stored as matrix-valued frames and exported in CSV format for downstream processing.

Frame selection was based on gait cycles, each consisting of a stance phase and a swing phase (Figure 5). The pressure value at each time point was calculated as the sum of pressure readings across all sensing elements (kPa) as an overall loading indicator. In Figure 5, the green bracket denotes 10 gait cycles, blue markers indicate frames within the stance phase, and red markers denote the swing phase. As the plantar load is negligible during swing, the sensors yield near-zero readings. Therefore, to enable frame-wise classification of gait patterns and only the stance-phase frames of the 10 cycles were retained for analysis.

The final dataset comprised 6994 frames in 3 gait patterns: slow walking (2590 frames), fast walking (2162 frames), and uphill walking (2242 frames). These frames served as input to the proposed models for training and evaluation.

Matrix-form plantar pressure signals exported by the F-Scan system were converted into frame-wise instantaneous plantar pressure matrices using Python 3.10.12 (Python Software Foundation) to enable deep learning. Each frame-wise matrix was normalized to the range (0, 1) using min-max normalization and rendered directly as a grayscale intensity image, in which pixel intensity represents normalized pressure magnitude. For visualization (Figure 6A), we additionally rendered the same normalized matrices as pseudocolor pressure maps using the perceptually uniform viridis colormap. The viridis colormap was used for visualization only. The grayscale images were resized to 64×64 pixels to ensure consistency across samples, and Gaussian noise was injected as a lightweight augmentation to improve noise robustness and mitigate overfitting. After preprocessing, the resulting input tensor for model training and evaluation had the shape (6994, 64, 64, 1).

In this study, 3 deep learning architectures were developed to address plantar pressure image classification, each designed with distinct structural characteristics. The first was a lightweight CNN (Light CNN), which served as the baseline model. It was constructed with 3 convolutional layers, each followed by max-pooling to progressively reduce dimensionality while preserving salient spatial features. Batch normalization was incorporated to stabilize the training process, while dropout layers were used to mitigate overfitting. A fully connected layer and a final softmax classifier were used to output the probabilities of the 3 gait categories, namely, slow walking, fast walking, and uphill walking. The structural design of this network is summarized in Table 2, which establishes the baseline framework for subsequent comparisons.

The second model was the AE-CNN cascade, which modularized the integration of an AE and a CNN classifier. In this design, the encoder of the AE first compressed the high-dimensional plantar pressure matrices into a compact latent representation, while the decoder simultaneously ensured that essential structural information could be reconstructed. The compressed latent features were then transferred to a CNN classifier for further convolutional processing and final classification. This staged cascade design effectively separated feature compression from classification, improving both the stability and the interpretability of the learning process. The encoder architecture used in the AE-CNN cascade is presented in Table 3, illustrating how feature reduction and classification were sequentially integrated.

To evaluate generalization while reducing the risk of information leakage from correlated frame-wise samples, data partitioning was performed at the participant level. Participants (n=13) were first split into a development set (n=10, 76.9%) and an independent held-out test set (n=3, 23.1%). Within the development set, a participant-wise grouped 9-fold cross-validation procedure was applied for model selection and stability assessment. In each fold, all frame-wise samples from participants in the training folds were used for model fitting, and all samples from participants in the held-out fold were used for validation. As 10 is not evenly divisible by 9, fold sizes differed by at most 1 participant (8 folds contained 1 participant and 1 fold contained 2 participants). After cross-validation, the final model configuration was retrained on the full development set and evaluated once on the held-out test set. Performance was reported using accuracy, precision, recall, and F₁-score.

The third architecture was the encoder-augmented CNN, which further extended the integration of AE and CNN by embedding the encoder directly into the CNN pipeline. Unlike the cascade structure, this hybrid design adopted an end-to-end framework, where the encoder served as the initial feature extractor and its outputs were directly connected to the subsequent CNN layers. This approach allowed the encoder and CNN to be jointly optimized, combining compact representation learning with the discriminative power of deep convolutional layers. The architectural layout of this model is summarized in Table 4, highlighting its streamlined structure and enhanced learning efficiency.

Model performance was assessed using accuracy, precision, recall (sensitivity), F₁-score, and confusion matrix:

All metrics were computed on the designated evaluation split without using any evaluation data for parameter updating. Unless stated otherwise, results are reported at the overall level to facilitate comparison among the baseline CNN, the AE-CNN cascade, the encoder-augmented CNN, and traditional classifiers; confusion matrices are additionally provided to illustrate class-wise error patterns. To reduce optimistic bias due to correlated frame-wise samples, all data splitting was performed at the participant level. Specifically, all stance-phase frames from the same participant were assigned to a single fold during cross-validation, ensuring that no participant’s frames appeared in both the training and validation sets within any iteration. Model performance was summarized using accuracy, precision, recall, and F₁-score.

This model development and evaluation study was reported with reference to the CREMLS (Consolidated Reporting of Machine Learning Studies) checklist. The completed checklist is provided in Multimedia Appendix 1.

This section summarizes the main comparative findings from the deep learning and classical machine learning models evaluated. Detailed training curves, optimization experiments, and hyperparameter tuning results are provided in Multimedia Appendix 2.

Among the baseline deep learning models, the Light CNN achieved an F₁-score of 94.44% on the held-out test set, whereas the AE-CNN cascade achieved an F₁-score of 92.45%.

Additional optimization analyses of the encoder-augmented CNN, including comparisons of downsampling strategies, batch normalization configurations, and bottleneck layer inclusion, supported the final model design reported here (Multimedia Appendix 2).

For the classical machine learning models trained on AE-derived features, support vector machine (SVM) with a radial basis function kernel performed best (F₁-score=93.76%), followed by k-nearest neighbors (F₁-score=91.73%) and random forest (F₁-score=88.54%).

The primary finding of this study is that the proposed encoder-augmented CNN architecture achieves superior performance (F₁-score=96.20%) in classifying dynamic gait patterns from plantar pressure images compared to both baseline deep learning models and classical machine learning classifiers (Figure 7; Table 5). This result supported the hypothesis that an integrated deep learning architecture, combining the feature extraction and dimensionality reduction capabilities of an AE with the classification power of a CNN, is highly effective for this task.

This model development and evaluation study demonstrates that the integration of AE-based representation learning with CNN architectures can achieve accurate recognition of gait patterns based on plantar pressure in a pilot dataset of healthy participants. Future work should validate generalization across broader demographics and clinical populations, evaluate robustness to real-world variability (eg, footwear, sensor noise, and speed fluctuations), and further strengthen interpretability and deployment feasibility for wearable or embedded applications. The performance of the Light CNN (F₁-score=94.44%) also demonstrated that even a lightweight, stand-alone CNN can effectively learn discriminative features from plantar pressure data. Among the classical methods, the support vector machine–RBF classifier proved to be the most robust, outperforming k-nearest neighbors and random forest, which aligns with its known strengths in handling high-dimensional feature spaces.

The study also highlights the importance of architectural choices in model optimization. Supplementary analyses of downsampling strategy, batch normalization placement, and bottleneck layer inclusion are provided in Multimedia Appendix 2. These experiments supported the final selection of the encoder-augmented CNN configuration reported in the main manuscript.

A notable and consistent finding in multiple models was the confusion between “fast walking” and “uphill walking” gaits (Figure 8). The encoder-augmented CNN, for example, misclassified 23 “uphill” instances as “fast walking.” This suggests that at similar walking speeds, the plantar pressure distributions for these 2 activities share substantial similarities. The primary differentiator may lie in subtle temporal features or pressure shifts related to gravitational resistance during uphill walking, which current spatial feature–focused models may not fully capture. In contrast, “slow walking” was classified with very high precision, indicating that variations in walking speed produce more distinct plantar pressure patterns than variations in surface uphill.

The use of deep learning, particularly CNNs, for plantar pressure analysis is consistent with recent trends in biomechanics and clinical research. Although many studies have successfully used CNNs for static pressure images (eg, for disease diagnosis), this research extends their application to dynamic gait classification. The accuracy achieved by the encoder-augmented CNN (96.21%) is competitive with or exceeds that reported in other studies using different sensor modalities or classification algorithms for similar tasks. The finding that an integrated AE-CNN architecture outperforms a standard CNN suggests that explicit feature learning and dimensionality reduction prior to classification can be a beneficial strategy for complex, high-variance data such as plantar pressure sequences. Clinical and translational implications should be interpreted with caution. This pilot study included only healthy young adults under controlled treadmill conditions and did not include participants with pathological gaits (eg, diabetes-related gait alteration). Therefore, while plantar pressure imaging is clinically relevant and the proposed framework shows technical promise, clinical screening performance has not been empirically validated here and requires future studies in clinical cohorts and real-world environments. In addition, our current approach performs frame-wise classification of stance-phase plantar pressure maps and does not explicitly model temporal dynamics across gait cycles; future work could incorporate sequence models (eg, temporal CNNs or long short-term memory) to better capture temporal gait signatures.

Although this study has yielded promising results, several limitations should be considered:

1. Single data source—plantar pressure data in this study were collected from a limited number of participants in a controlled laboratory environment. Whether the model’s generalization ability can be extended to populations with different ages, genders, weights, or specific pathological gaits (eg, flat feet and diabetic foot) requires further validation.

2. Model interpretability—compared to classical machine learning models, deep learning models are often regarded as “black boxes,” with less transparent decision-making processes. Although this study validated the effectiveness of the model, it did not explore which specific regions or features of plantar pressure images the model used for classification.

3. Limited gait types—the study only covered 3 specific dynamic gaits. The applicability of the model to more complex daily activities, such as walking downhill, turning, or climbing stairs, has yet to be determined.

In addition, the evaluation in this study was reported primarily at the frame level. Although frame-wise metrics are useful for model comparison, participant-level performance (eg, aggregating frame-level predictions to participant-level decisions) should be reported in future studies to better reflect real-world use. Moreover, statistical uncertainty (eg, CIs via bootstrapping) and systematic robustness tests under controlled perturbations (eg, varying noise intensity or sensor shift) were not performed in this study and remain important directions for future work.

This study developed and evaluated a deep learning architecture named encoder-augmented CNN for gait classification using plantar pressure images. The model combines the feature extraction capabilities of an AE with the classification strengths of a CNN. Through systematic structural optimization, it ultimately achieved an accuracy of 96.21% in the 3-class dynamic-gait recognition task, outperforming classical machine learning methods and other deep learning variants in this study.

On the basis of the findings and limitations of this research, future studies could proceed in the following directions: