3D Digital Twin Visualization of Multiclass GRF-Based Gait Disorder Classification

# 3D Digital Twin Visualization of Multiclass GRF-Based Gait Disorder Classification
Source: [https://arxiv.org/html/2609.12442](https://arxiv.org/html/2609.12442)
Nayoung SonAffiliation:Department of Software, Yonsei University, Wonju, 26493, Republic of KoreaE\-mail[mshin@yonsei\.ac\.kr](mailto:[email protected])Minwoo Shin††thanks:Corresponding author\.Affiliation:Department of Software, Yonsei University, Wonju, 26493, Republic of KoreaE\-mail[mshin@yonsei\.ac\.kr](mailto:[email protected])

###### Abstract

Automated gait analysis requires accurate classification and interpretable outputs\. We propose an integrated framework for classifying healthy gait and multiple musculoskeletal impairment groups using bilateral ground reaction force \(GRF\) and center\-of\-pressure \(COP\) signals\. The signals were normalized over the stance phase and standardized using training\-set statistics\. The model achieved a validation accuracy of 99\.00% and a test accuracy of 90\.07% under a session\-level split\. Class\-specificϵ\\epsilon\-LRP identified positive and negative contributions across both sides, multiple signal components, and different stance phases\. Separately, the processed GRF signals and model predictions were synchronized within a Blender\-based 3D visualization, enabling sample\-level inspection of gait trials and classification results\. The proposed framework integrates classification, explainability, and 3D visualization to improve model transparency\. The source code is available in the following repository:[https://github\.com/nyoico/grf\-gait\-3d\-visualization\.git](https://github.com/nyoico/grf-gait-3d-visualization.git)

###### Keywords:

Gait Disorder Ground Reaction Force Multiclass Classification Transformer Explainable AI 3D Visualization

## 1Introduction

Instrumented gait analysis provides objective measurements for assessing human locomotion and identifying pathological gait patterns\[[2](https://arxiv.org/html/2609.12442#bib.bib1),[26](https://arxiv.org/html/2609.12442#bib.bib2),[19](https://arxiv.org/html/2609.12442#bib.bib3)\]\. Among biomechanical signals, ground reaction force \(GRF\) is widely used because it reflects weight acceptance, propulsion, balance, and limb loading during walking\[[12](https://arxiv.org/html/2609.12442#bib.bib7),[20](https://arxiv.org/html/2609.12442#bib.bib9)\]\. However, clinical GRF recordings are high\-dimensional, temporally dependent, and variable across subjects and impairment types, making manual interpretation difficult and motivating automated gait classification systems\[[5](https://arxiv.org/html/2609.12442#bib.bib10),[6](https://arxiv.org/html/2609.12442#bib.bib11),[9](https://arxiv.org/html/2609.12442#bib.bib12),[17](https://arxiv.org/html/2609.12442#bib.bib5),[24](https://arxiv.org/html/2609.12442#bib.bib26),[7](https://arxiv.org/html/2609.12442#bib.bib27)\]\.

Machine learning has been widely explored for gait disorder classification using GRF and related biomechanical measurements\[[9](https://arxiv.org/html/2609.12442#bib.bib12),[21](https://arxiv.org/html/2609.12442#bib.bib4),[17](https://arxiv.org/html/2609.12442#bib.bib5),[11](https://arxiv.org/html/2609.12442#bib.bib6)\]\. Conventional approaches, including support vector machines, nearest\-neighbor classifiers, self\-organizing neural networks, and other analytical models, have been used to categorize gait patterns into clinically meaningful groups\[[4](https://arxiv.org/html/2609.12442#bib.bib13),[18](https://arxiv.org/html/2609.12442#bib.bib14),[3](https://arxiv.org/html/2609.12442#bib.bib15)\]\. Large\-scale annotated datasets such as GaitRec have further enabled data\-driven modeling of healthy and impaired gait, including musculoskeletal impairments involving the hip, knee, ankle, and calcaneus\[[12](https://arxiv.org/html/2609.12442#bib.bib7)\]\. Related public GRF resources, such as the Gutenberg Gait Database, also support reproducible gait modeling and benchmarking\[[13](https://arxiv.org/html/2609.12442#bib.bib8)\]\. Nevertheless, two limitations remain important\. First, many classification models provide limited interpretability, making it difficult to understand which gait phases or force components drive the decision\[[8](https://arxiv.org/html/2609.12442#bib.bib16)\]\. Second, model outputs are often presented only as numerical metrics or plots, limiting intuitive inspection of individual gait trials; visual analytics and digital twin approaches provide a complementary route for sample\-level interpretation\[[22](https://arxiv.org/html/2609.12442#bib.bib17),[16](https://arxiv.org/html/2609.12442#bib.bib21),[25](https://arxiv.org/html/2609.12442#bib.bib22)\]\.

To address these limitations, we propose an integrated framework for interpretable bilateral gait classification and digital twin\-based visualization\. The proposed model uses an encoder–decoder Transformer architecture to classify musculoskeletal gait patterns from left and right GRF/COP signals, motivated by the success of attention\-based models in sequence and multivariate time\-series representation learning\[[14](https://arxiv.org/html/2609.12442#bib.bib18)\]\. To improve transparency, we apply class\-specificϵ\\epsilon\-LRP to identify temporally localized and channel\-wise relevance patterns across the stance phase\[[1](https://arxiv.org/html/2609.12442#bib.bib19),[23](https://arxiv.org/html/2609.12442#bib.bib20)\]\. In addition, we develop a Blender\-based gait digital twin simulation that synchronizes GRF values, ground\-truth labels, and predicted labels within a 3D virtual gait environment\[[27](https://arxiv.org/html/2609.12442#bib.bib25),[10](https://arxiv.org/html/2609.12442#bib.bib23),[15](https://arxiv.org/html/2609.12442#bib.bib24)\]\. This enables both quantitative evaluation and sample\-level qualitative inspection of model predictions\.

The main contributions of this study are as follows\. First, we develop a bilateral Transformer\-based classifier for multiclass musculoskeletal gait classification using time\-normalized GRF/COP signals\. Second, we introduce class\-specificϵ\\epsilon\-LRP analysis to visualize positive and negative relevance across stance phases and force components\. Third, we integrate the model outputs into a Blender\-based data\-driven gait digital twin, allowing individual test samples to be examined together with synchronized biomechanical signals and classification results\.

## 2Method

### 2\.1Data and Input Definition

Raw GRF and COP data were preprocessed at the session and trial level\. Each gait trial was represented as 101 time\-normalized stance\-phase points\. For each side, five channels—COP\_AP, COP\_ML, F\_AP, F\_ML, and F\_V—were aligned using SESSION\_ID and TRIAL\_ID, and only trials common to all channels were retained\. Here, COP denotes the center of pressure, while AP, ML, and V indicate the anterior–posterior, mediolateral, and vertical directions, respectively\. Therefore, COP\_AP and COP\_ML represent the center\-of\-pressure trajectories in the anterior–posterior and mediolateral directions, while F\_AP, F\_ML, and F\_V represent the ground reaction force components along the anterior–posterior, mediolateral, and vertical axes\.

SESSION\_ID denotes the unique identifier of each recording session, and TRIAL\_ID indicates the repeated trial index within a session\. Diagnostic labels were assigned from CLASS\_LABEL\_DETAILED metadata using SESSION\_ID, where CLASS\_LABEL\_DETAILED represents the detailed diagnostic class of each session\. The classifier used 12 output labels, including healthy controls and musculoskeletal impairment groups\. Table[1](https://arxiv.org/html/2609.12442#S2.T1)summarizes the class definitions and their use across the data splits\. Because the C\_A class contained only one sample in the entire dataset, it was retained only in the training set and excluded from validation and test evaluation\. For the remaining 11 classes, a stratified split was performed at the session level to prevent data leakage among the training, validation, and test sets\. Left and right side data were saved separately as arrays with shapeN×101×5N\\times 101\\times 5, whereNNdenotes the number of gait trials\.

Table 1:Classification labels and evaluation usage\.
### 2\.2Network

The proposed model uses an encoder–decoder Transformer for bilateral gait classification, as illustrated in Fig\.[1](https://arxiv.org/html/2609.12442#S2.F1)\. Left and right GRF/COP sequences are first projected using side\-specific 1D convolutional embeddings and combined with sinusoidal positional encoding, thereby preserving bilateral input characteristics before temporal modeling\. The Transformer encoder models temporal dependencies across the 101\-point stance phase using multi\-head self\-attention, while a learned decoder query attends to the encoded gait representation through cross\-attention\. As shown in Fig\.[1](https://arxiv.org/html/2609.12442#S2.F1), the resulting decoder feature is passed to dropout and a linear classifier to produce 12 class logits\. The main architectural settings are summarized in Table[2](https://arxiv.org/html/2609.12442#S3.T2)\.

![Refer to caption](https://arxiv.org/html/2609.12442v1/figure/architecture_lr.png)Figure 1:Overall architecture of the proposed Transformer encoder–decoder classifier\. Bilateral COP and GRF sequences are processed through side\-specific convolutional embeddings and positional encoding\. The encoder models temporal gait representations, and a learned decoder query attends to the encoder memory to produce 12 class logits\.
### 2\.3Architecture\-Awareϵ\\epsilon\-LRP for Gait Model Explanation

Class\-specific relevance propagation was performed using each output logit as the initial relevance score\. Relevance was propagated backward through the classifier, decoder, encoder, and convolutional embedding layers to obtain time\- and channel\-wise relevance scores for signals from both sides\. For class\-level analysis, relevance scores from all test samples within the same class were averaged to reduce trial\-level variability\.

The implemented method uses customizedϵ\\epsilon\-LRP rules adapted to the proposed architecture\. Epsilon\-stabilized rules were applied to the linear classifier and convolutional layers, feed\-forward networks were handled through their Conv1D–ReLU–Conv1D structure, and attention modules were treated with a pass\-through strategy\. Because the left and right embeddings are combined by summation, relevance was equally distributed to both branches and propagated back to the input channels\. Signed relevance was visualized over ML, AP, and vertical force components: positive relevance indicates evidence supporting the target class logit, whereas negative relevance indicates evidence suppressing it or separating it from competing classes\. These relevance maps were interpreted as model\-derived evidence rather than direct clinical causality\.

### 2\.43D Gait Visualization in Blender

A Blender\-based visualization module was developed to inspect individual test trials in a virtual gait analysis scene\. The module reads the inference CSV containing true labels, predicted labels, and bilateral GRF time series, then updates the displayed GRF values according to the current stance\-phase frame\. The scene includes a humanoid model, bilateral force plates, and a text panel showing the sample index, time index, GRF components, true class, and predicted class\. This provides a sample\-level interface for jointly reviewing gait signals and model predictions\.

![Refer to caption](https://arxiv.org/html/2609.12442v1/figure/blender/Blender_image.png)Figure 2:Blender\-based 3D visualization of the gait classification result\. A humanoid model is placed on bilateral force plates, while the text panel displays the current sample index, stance\-phase time index, bilateral GRF values, true class, and predicted class\.

## 3Experiments

### 3\.1Dataset and Evaluation Protocol

Experiments were conducted on the post\-processed GaitRec dataset using the 12 target classes in Table[1](https://arxiv.org/html/2609.12442#S2.T1)\. The session\-level split produced 61,266 training, 1,894 validation, and 1,994 independent test samples\. Table[3](https://arxiv.org/html/2609.12442#S3.T3)summarizes the training hyperparameters and inference configuration used in the experiments, including the data split setting, optimization parameters, learning\-rate scheduler, early stopping criterion, and latency measurement protocol\. The model was trained using cross\-entropy loss with label smoothing, AdamW optimization, ReduceLROnPlateau scheduling, and early stopping based on validation performance\. The best validation checkpoint was used for final testing\. Performance was evaluated using accuracy, confusion matrix, class\-wise precision/recall/F1\-score, and single\-sample inference latency after warm\-up iterations\.

Table 2:Model configuration of the proposed Transformer\-based gait classification network\.Table 3:Training hyperparameters and inference configuration\. Latency was measured on CUDA with batch size 1 after 20 warm\-up iterations and averaged over 100 repeats\.
### 3\.2Quantitative Classification Results

As summarized in Table[4](https://arxiv.org/html/2609.12442#S3.T4), the model achieved 98\.997% validation accuracy at epoch 97 and 90\.07% test accuracy on 1,994 independent test samples\.

Table[5](https://arxiv.org/html/2609.12442#S3.T5)reports the class\-wise recall on the independent test set\. High recall was observed for several classes, including C\_F, H\_F, K\_F, and HC, whereas relatively lower recall was observed for K\_R and A\_R\. These results suggest that the proposed model performed well for most gait classes, while some clinically related impairment groups remained more difficult to distinguish due to similar biomechanical characteristics\. C\_A was not evaluated because it had no test samples\. The final model contained 7\.38M trainable parameters and achieved 4\.89±\\pm0\.36 ms single\-sample inference latency on an RTX 5080\.

Table 4:Quantitative summary of the proposed model\.Table 5:Class\-wise recall \(%\) on the independent test set\. Support and correct denote total and correctly classified samples\. C\_A was excluded because it was absent from the test split\.
### 3\.3Class\-Specificϵ\\epsilon\-LRP Analysis

Class\-specificϵ\\epsilon\-LRP was applied to the 11 classes present in the test split\. Fig\.[3](https://arxiv.org/html/2609.12442#S3.F3)shows the class\-specific mean signedϵ\\epsilon\-LRP maps for the left and right sides\. Each subplot visualizes the averaged GRF waveform together with relevance bands across the stance phase, where red bands indicate positive relevance supporting the corresponding class prediction and blue bands indicate negative relevance opposing the prediction\. The relevance maps showed class\- and side\-dependent evidence across stance phases and across ML, AP, and vertical force components\. In particular, the model did not rely uniformly on the entire stance phase; instead, it emphasized localized temporal regions that differed by class and side\. This suggests that the proposed Transformer captured discriminative gait features related to side\-specific loading, braking/propulsion, and vertical force patterns\. Signed relevance indicated both supporting and opposing evidence, showing that the model used localized biomechanical patterns to separate impairment classes\.

![Refer to caption](https://arxiv.org/html/2609.12442v1/figure/Fig3_lr.png)Figure 3:Class\-specific mean signedϵ\\epsilon\-LRP maps for the 11 gait classes represented in the test set\. Red and blue bands indicate positive and negative relevance, respectively, across the stance phase for the left\- and right\-side ML, AP, and vertical force components\.
### 3\.4Blender\-Based Gait Digital Twin Simulation

The Blender\-based digital twin was used for qualitative inspection of individual test samples\. As shown in Fig\.[2](https://arxiv.org/html/2609.12442#S2.F2), the scene displays a humanoid model on bilateral force plates and an information panel with synchronized GRF values, true labels, and predicted labels\. This interface complements quantitative metrics andϵ\\epsilon\-LRP by enabling sample\-specific review of gait signals and model predictions\.

## 4Conclusion

This study proposed an integrated framework for gait classification, model interpretation, and 3D digital twin visualization using bilateral GRF/COP signals\. The encoder–decoder Transformer achieved 90\.07% test accuracy with fast single\-sample inference, while class\-specificϵ\\epsilon\-LRP revealed temporally localized relevance patterns across force components and sides\. The Blender\-based digital twin enabled sample\-level inspection with synchronized GRF values, true labels, and predicted labels\. Future work will incorporate richer motion dynamics, external validation, and broader clinical evaluation\.

#### Acknowledgements

This work was supported by the Korea Medical Device Development Foundation grant funded by the Korean government \(the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health and Welfare, and the Ministry of Food and Drug Safety\) \(Grant No\. RS\-2026\-25543484\)\.

This research was also supported by the ANCHOR Program through the Gangwon ANCHOR Center, funded by the Ministry of Education \(MOE\) and Gangwon State \(G\.S\.\), Republic of Korea \(Grant No\. 2026\-ANCHOR\-10\-006\)\.

This research was further supported by the Ministry of Science and ICT \(MSIT\), Korea, under the National Program in Medical AI Semiconductor \(Grant No\. 2024\-0\-00096\), supervised by the Institute of Information & Communications Technology Planning & Evaluation \(IITP\) in 2026\.

#### Disclosure of Interests\.

The authors have no competing interests to declare that are relevant to the content of this article\.

## References

- \[1\]\(2015\)On pixel\-wise explanations for non\-linear classifier decisions by layer\-wise relevance propagation\.PLOS ONE10\(7\),pp\. e0130140\.External Links:[Document](https://dx.doi.org/10.1371/journal.pone.0130140)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p3.1)\.
- \[2\]R\. Baker\(2006\)Gait analysis methods in rehabilitation\.Journal of NeuroEngineering and Rehabilitation3,pp\. 4\.External Links:[Document](https://dx.doi.org/10.1186/1743-0003-3-4)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1)\.
- \[3\]G\. J\. Barton, P\. J\. G\. Lisboa, A\. Lees, and S\. Attfield\(2007\)Gait quality assessment using self\-organising artificial neural networks\.Gait & Posture25\(3\),pp\. 374–379\.External Links:[Document](https://dx.doi.org/10.1016/j.gaitpost.2006.05.003)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[4\]R\. Begg and J\. Kamruzzaman\(2005\)A machine learning approach for automated recognition of movement patterns using basic, kinetic and kinematic gait data\.Journal of Biomechanics38\(3\),pp\. 401–408\.External Links:[Document](https://dx.doi.org/10.1016/j.jbiomech.2004.05.002)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[5\]T\. Chau\(2001\)A review of analytical techniques for gait data\. part 1: fuzzy, statistical and fractal methods\.Gait & Posture13\(1\),pp\. 49–66\.External Links:[Document](https://dx.doi.org/10.1016/S0966-6362%2800%2900094-1)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1)\.
- \[6\]T\. Chau\(2001\)A review of analytical techniques for gait data\. part 2: neural network and wavelet methods\.Gait & Posture13\(2\),pp\. 102–120\.External Links:[Document](https://dx.doi.org/10.1016/S0966-6362%2800%2900095-3)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1)\.
- \[7\]S\. Cho, M\. Seo, M\. Shin, and K\. Yoon\(2026\)Towards digital twin of RF ablation: real\-time prediction of time\-dependent thermal effects using transformer\.InDigital Twin for Healthcare,L\. Li, V\. Jirsa, J\. Feng, J\. Deng, L\. Dede’, S\. An, Y\. Lyu, and X\. Liu \(Eds\.\),Cham,pp\. 69–78\.External Links:ISBN 978\-3\-032\-07694\-6Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1)\.
- \[8\]C\. Dindorf, W\. Teufl, B\. Taetz, G\. Bleser, and M\. Frohlich\(2020\)Interpretability of input representations for gait classification in patients after total hip arthroplasty\.Sensors20\(16\),pp\. 4385\.External Links:[Document](https://dx.doi.org/10.3390/s20164385)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[9\]J\. Figueiredo, C\. P\. Santos, and J\. C\. Moreno\(2018\)Automatic recognition of gait patterns in human motor disorders using machine learning: a review\.Medical Engineering & Physics53,pp\. 1–12\.External Links:[Document](https://dx.doi.org/10.1016/j.medengphy.2017.12.006)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1),[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[10\]A\. Fuller, Z\. Fan, C\. Day, and C\. Barlow\(2020\)Digital twin: enabling technologies, challenges and open research\.IEEE Access8,pp\. 108952–108971\.External Links:[Document](https://dx.doi.org/10.1109/ACCESS.2020.2998358)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p3.1)\.
- \[11\]E\. J\. Harris, I\. Khoo, and E\. Demircan\(2022\)A survey of human gait\-based artificial intelligence applications\.Frontiers in Robotics and AI8,pp\. 749274\.External Links:[Document](https://dx.doi.org/10.3389/frobt.2021.749274)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[12\]B\. Horsak, D\. Slijepcevic, A\. Raberger, C\. Schwab, M\. Worisch, and M\. Zeppelzauer\(2020\)GaitRec, a large\-scale ground reaction force dataset of healthy and impaired gait\.Scientific Data7,pp\. 143\.External Links:[Document](https://dx.doi.org/10.1038/s41597-020-0481-z)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1),[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[13\]F\. Horst, D\. Slijepcevic, M\. Simak, and W\. I\. Schollhorn\(2021\)Gutenberg gait database, a ground reaction force database of level overground walking in healthy individuals\.Scientific Data8,pp\. 232\.External Links:[Document](https://dx.doi.org/10.1038/s41597-021-01014-6)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[14\]H\. Ismail Fawaz, G\. Forestier, J\. Weber, L\. Idoumghar, and P\. Muller\(2019\)Deep learning for time series classification: a review\.Data Mining and Knowledge Discovery33\(4\),pp\. 917–963\.External Links:[Document](https://dx.doi.org/10.1007/s10618-019-00619-1)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p3.1)\.
- \[15\]D\. Jones, C\. Snider, A\. Nassehi, J\. Yon, and B\. Hicks\(2020\)Characterising the digital twin: a systematic literature review\.CIRP Journal of Manufacturing Science and Technology29,pp\. 36–52\.External Links:[Document](https://dx.doi.org/10.1016/j.cirpj.2020.02.002)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p3.1)\.
- \[16\]E\. Katsoulakis, Q\. Wang, H\. Wu, L\. Shahriyari, R\. Fletcher, J\. Liu, L\. Achenie, H\. Liu, P\. Jackson, Y\. Xiao, T\. Syeda\-Mahmood, R\. Tuli, and J\. Deng\(2024\)Digital twins for health: a scoping review\.npj Digital Medicine7,pp\. 77\.External Links:[Document](https://dx.doi.org/10.1038/s41746-024-01073-0)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[17\]P\. Khera and N\. Kumar\(2020\)Role of machine learning in gait analysis: a review\.Journal of Medical Engineering & Technology44\(8\),pp\. 441–467\.External Links:[Document](https://dx.doi.org/10.1080/03091902.2020.1822940)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1),[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[18\]N\. Mezghani, S\. Husse, K\. Boivin, K\. Turcot, R\. Aissaoui, N\. Hagemeister, and J\. A\. de Guise\(2008\)Automatic classification of asymptomatic and osteoarthritis knee gait patterns using kinematic data features and the nearest neighbor classifier\.IEEE Transactions on Biomedical Engineering55\(3\),pp\. 1230–1232\.External Links:[Document](https://dx.doi.org/10.1109/TBME.2007.905388)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[19\]A\. Muro\-de\-la\-Herran, B\. Garcia\-Zapirain, and A\. Mendez\-Zorrilla\(2014\)Gait analysis methods: an overview of wearable and non\-wearable systems, highlighting clinical applications\.Sensors14\(2\),pp\. 3362–3394\.External Links:[Document](https://dx.doi.org/10.3390/s140203362)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1)\.
- \[20\]J\. S\. Park and C\. H\. Kim\(2022\)Ground\-reaction\-force\-based gait analysis and its application to gait disorder assessment: new indices for quantifying walking behavior\.Sensors22\(19\),pp\. 7558\.External Links:[Document](https://dx.doi.org/10.3390/s22197558)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1)\.
- \[21\]C\. Prakash, R\. Kumar, and N\. Mittal\(2018\)Recent developments in human gait research: parameters, approaches, applications, machine learning techniques, datasets and challenges\.Artificial Intelligence Review49\(1\),pp\. 1–40\.External Links:[Document](https://dx.doi.org/10.1007/s10462-016-9514-6)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[22\]A\. Rind, D\. Slijepcevic, M\. Zeppelzauer, F\. Unglaube, A\. Kranzl, and B\. Horsak\(2022\)Trustworthy visual analytics in clinical gait analysis: a case study for patients with cerebral palsy\.In2022 IEEE Workshop on TRust and EXpertise in Visual Analytics \(TREX\),pp\. 8–15\.External Links:[Document](https://dx.doi.org/10.1109/TREX57753.2022.00006)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[23\]W\. Samek, A\. Binder, G\. Montavon, S\. Lapuschkin, and K\. Muller\(2017\)Evaluating the visualization of what a deep neural network has learned\.IEEE Transactions on Neural Networks and Learning Systems28\(11\),pp\. 2660–2673\.External Links:[Document](https://dx.doi.org/10.1109/TNNLS.2016.2599820)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p3.1)\.
- \[24\]M\. Seo, M\. Shin, G\. Noh, S\. Yoo, and K\. Yoon\(2026\)Acoustic simulation with deep learning for low\-intensity transcranial focused ultrasound digital twins\.InDigital Twin for Healthcare,L\. Li, V\. Jirsa, J\. Feng, J\. Deng, L\. Dede’, S\. An, Y\. Lyu, and X\. Liu \(Eds\.\),Cham,pp\. 58–68\.External Links:ISBN 978\-3\-032\-07694\-6Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1)\.
- \[25\]L\. Uhlenberg, A\. Derungs, and O\. Amft\(2023\)Co\-simulation of human digital twins and wearable inertial sensors to analyse gait event estimation\.Frontiers in Bioengineering and Biotechnology11,pp\. 1104000\.External Links:[Document](https://dx.doi.org/10.3389/fbioe.2023.1104000)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p2.1)\.
- \[26\]M\. W\. Whittle\(1996\)Clinical gait analysis: a review\.Human Movement Science15\(3\),pp\. 369–387\.External Links:[Document](https://dx.doi.org/10.1016/0167-9457%2896%2900006-1)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p1.1)\.
- \[27\]R\. D\. Wickes\(2009\)Blender overview\.InFoundation Blender Compositing,pp\. 1–23\.External Links:[Document](https://dx.doi.org/10.1007/978-1-4302-1977-4%5F1)Cited by:[§1](https://arxiv.org/html/2609.12442#S1.p3.1)\.