Abstract
Background
Understanding joint loading and the crucial role of joint moments is essential for developing treatment strategies in gait analysis, which often requires the precise estimation of joint moments through an inverse dynamic approach. This process necessitates the use of a force plate synchronized with a motion capture system. However, effectively capturing ground reaction force in typically developing (TD) children and those with congenital talipes equino varus (CTEV) presents challenges, while the availability and high cost of additional force plates pose additional challenges. Therefore the study aimed to develop, train, and identify the most effective machine learning (ML) model to predict joint moments from kinematics for TD children and those with CTEV.
Method
In a study at the Gait Lab, 13 children with bilateral CTEV and 17 TD children underwent gait analysis to measure kinematics and kinetics, using a 12-camera Qualisys Motion Capture System and an AMTI force plate. ML models were then trained to predict joint moments from kinematic data as input.
Results
The random forest regressor and deep neural networks (DNN) proved most effective in predicting joint moments from kinematics for TD children, yielding better results. The Random Forest regressor achieved an average r of 0.75 and nRMSE of 23.03 % for TD children, and r of 0.74 and 23.82 % for CTEV. DNN achieved an average r of 0.75 and nRMSE of 22.83 % for TD children, and r of 0.76 and nRMSE of 23.9 % for CTEV.
Conclusions
The findings suggest that using machine learning to predict joint moments from kinematics shows moderate potential as an alternative to traditional gait analysis methods for both TD children and those with CTEV. Despite its potential, the current prediction accuracy limitations hinder the immediate clinical application of these techniques for decision-making in a pediatric population.
Keywords: CTEV, Machine learning, Gait analysis, Joint moments prediction, Gait kinematics
1. Introduction
Assessing joint health during gait necessitates an understanding of joint loading.1 Joint contact forces offer crucial insights into the loads affecting joints, which are key to assessing joint health. However, estimating these forces requires a sophisticated musculoskeletal model, adding complexity to the process.2, 3, 4 Consequently, joint moments are often used as a surrogate for joint contact forces.5, 6, 7 The joint moment is also crucial for developing effective treatment strategies, including orthotic devices, physical therapy, or surgical interventions, to improve mobility, reduce pain, and prevent long-term complications. In a typical gait lab equipped with a motion capture system, kinematics, such as joint angles during gait, are computed using a camera system and markers. To estimate joint moments using an inverse dynamic approach, an additional force plate synchronized with the motion capture system is indispensable. Capturing ground reaction force (GRF) during natural walking, especially in typically developing (TD) children, can be challenging. Moreover, it is particularly arduous to capture GRF in cases such as Congenital Talipes Equino Varus (CTEV). Furthermore, an additional force plate is not always readily available and can be costly to obtain. CTEV stands out as a prevalent congenital anomaly of the foot, distinguished by hindfoot varus and equinus, forefoot adduction, and midfoot cavus. The Ponseti method, a treatment widely embraced across the globe, is employed to rectify this deformity soon after a child is born. Nonetheless, there's an ongoing interest in exploring biomechanical changes in children who have received Ponseti treatment.
To address the challenges associated with estimating joint moments, machine learning (ML) can serve as a valuable tool for predicting joint moments based on joint kinematics. In recent years, several attempts have been made to predict joint moments employing kinematic data as input in ML models for both adult individuals8 and those with cerebral palsy (CP),9 as well as for activities such as gait and single-leg squatting.10 A study devised an artificial neural network (ANN) to predict external knee moments during different locomotion tasks, utilizing data collected from two wearable sensors.11 Additionally, efforts have also been made to utilize a machine-learning approach to estimate joint contact forces.12,13 However, to the author's knowledge, there has been no investigation to date into predicting joint moments in TD and CTEV children using ML techniques.
The objective of the study was to develop, train, and identify the most effective machine learning models for predicting knee, hip, and ankle moments using kinematic inputs from TD children and those with CTEV. To achieve this goal, various ML algorithms were trained using kinematics to predict joint moments, and the accuracy of the predicted joint moments was subsequently calculated.
2. Materials and methods
2.1. Subjects
A study conducted in the Gait Lab involved 13 children (aged 6–15 years both males and females), totaling 26 legs, all with bilateral CTEV, recruited from the outpatient clinic. Inclusion criteria encompassed children diagnosed with idiopathic CTEV, aged above 6 years, who had completed Ponseti treatment and been off braces for at least 2 years. Exclusion criteria included children with neurologic clubfoot (such as AMC, CP, etc.), and those unable to cooperate for gait analysis evaluation. Additionally, the study included 17 typically developing children (aged 6–15 years, both males and females), totaling 34 legs. Ethical approval was obtained, and informed consent was secured from the parents/guardians of the participating children.
2.2. Gait analysis
Gait analysis was performed on TD children and those with CTEV to measure kinematics and kinetics. Data collection utilized a 12-camera Qualisys® 3D Motion Capture System operating at 300 Hz frequency (Qualisys AB, Götebor, Sweden). GRF was recorded at 1200 Hz using an AMTI force plate. All children were instructed to walk barefoot at their self-selected comfortable pace, replicating their daily gait. The collected data underwent initial processing with Qualisys Track Manager v2.15 (Qualisys AB, Götebor, Sweden) software. Subsequently, Visual 3D (C-Motion, Germantown, MD, USA) software was utilized to compute joint angle kinematics and kinetics. Due to the unavailability of GRF data during the swing phase, only stance phases were utilized, with directly calculated moment data interpolated to a standard length from 0 to 60 % of the Gait cycle.
2.3. Machine learning
Various ML techniques were employed to predict joint moments using joint kinematics, such as polynomial ridge regression, random forest regressor, gradient boosting regressor, and deep neural network (DNN). The dataset was divided into two parts: 1) The training dataset, comprising 80 %, was utilized for tuning the model. 2) The testing dataset, comprising 20 % was utilized to test the model. It's important to note that the testing datasets were entirely separate from the training and testing datasets. Model performance was evaluated using error metrics such as the coefficient of correlation (r), and normalized root mean squared error (nRMSE) that was normalized to the mean range of the experimental data.9 The machine learning analysis was performed in Python (Python Software Foundation, USA) using scikit-learn.14 Fig. 1 illustrates the workflow of the study.
Fig. 1.
Workflow of the study.
3. Results
Fig. 2 shows the joint moment predictions of various ML models for TD individuals, while Fig. 3 shows the predictions for CTEV groups, highlighting the model's ability to anticipate joint moments across varied patterns.
Fig. 2.
Comparison of joint moments predicted by different machine learning models with joint moment estimated by inverse dynamics analysis for TD children.The black line indicates the inverse dynamic joint moment, whereas the magenta, blue, red, and green lines represent the predicted joint moment from polynomial ridge regression, random forest regressor, gradient boosting regressor, and deep neural network, respectively, with shaded standard deviation.
Fig. 3.
Comparison of joint moments predicted by different machine learning models with joint moment estimated by inverse dynamics analysis for CTEV children. The black line indicates the inverse dynamic joint moment, whereas the magenta, blue, red, and green lines represent the predicted joint moment from polynomial ridge regression, random forest regressor, gradient boosting regressor, and deep neural network, respectively, with shaded standard deviation.
3.1. Polynomial ridge regression
Table 1 displays the r and nRMSE values achieved for the moments within both TD and CTEV groups using polynomial regression. In the TD group, superior predictions were observed for dorsi-plantar flexion and hip flexion-extension moments compared to hip abduction-adduction and knee flexion-extension moments. The model predicted joint moments for TD children with an average correlation r of 0.74 and a nRMSE of 31.77 %. Similarly, for CTEV, the mean r was 0.75 and the nRMSE was 28.41 %.
Table 1.
r, nRMSE values of predicted continuous moments using polynomial ridge regression.
| Moment | TD |
CTEV |
|||
|---|---|---|---|---|---|
| r | nRMSE % | r | nRMSE % | ||
| Hip | flexion-extension | 0.78 | 17.72 | 0.88 | 18.35 |
| adduction-abduction | 0.79 | 32.66 | 0.82 | 24.59 | |
| Knee | flexion-extension | 0.77 | 26.71 | 0.77 | 31.27 |
| varus-valgus | 0.57 | 32.46 | 0.43 | 39.69 | |
| Ankle | dorsi-plantar flexion | 0.81 | 38.01 | 0.84 | 24.64 |
| inversion-eversion | 0.75 | 43.07 | 0.78 | 31.93 | |
3.2. Random forest regressor
Table 2 displays the r and nRMSE values achieved for the moments within both TD and CTEV groups using random forest. The model predicted joint moments for TD children with an average correlation r of 0.75 and a nRMSE of 23.02 %. Similarly, for CTEV, the mean r was 0.74 and the nRMSE was 23.82 %.
Table 2.
r, nRMSE values of predicted continuous moments using random forest regressor.
| Moment | TD |
CTEV |
|||
|---|---|---|---|---|---|
| r | nRMSE % | r | nRMSE % | ||
| Hip | flexion-extension | 0.79 | 14.13 | 0.92 | 12.09 |
| adduction-abduction | 0.69 | 24.98 | 0.71 | 26.65 | |
| Knee | flexion-extension | 0.86 | 19.49 | 0.72 | 21.48 |
| varus-valgus | 0.51 | 30.35 | 0.46 | 31.22 | |
| Ankle | dorsi-plantar flexion | 0.87 | 23.76 | 0.86 | 19.64 |
| inversion - eversion | 0.83 | 25.42 | 0.79 | 31.88 | |
3.3. Gradient boosting regressor
Table 3 displays the r and nRMSE values achieved for the moments within both TD and CTEV groups using gradient boosting. The model predicted joint moments for TD children with an average correlation r of 0.69 and an nRMSE of 26.6 %. Similarly, for CTEV, the mean r was 0.78 and the nRMSE was 25.01 %. However, artifacts can be seen (Fig. 2) in predicted joint moments.
Table 3.
r, nRMSE values of predicted continuous moments using gradient boosting regressor.
| Moment | TD |
CTEV |
|||
|---|---|---|---|---|---|
| r | nRMSE % | r | nRMSE % | ||
| Hip | flexion-extension | 0.6 | 24.08 | 0.90 | 14.22 |
| adduction-abduction | 0.65 | 24.95 | 0.77 | 24.70 | |
| Knee | flexion-extension | 0.73 | 24.76 | 0.83 | 23.75 |
| varus-valgus | 0.48 | 32.5 | 0.51 | 33.27 | |
| Ankle | dorsi-plantar flexion | 0.86 | 26.35 | 0.82 | 22.94 |
| inversion - eversion | 0.84 | 26.92 | 0.85 | 31.19 | |
3.4. Deep neural network
Table 4 displays the r and nRMSE values achieved for the moments within both TD and CTEV groups using a deep neural network. Table 4 displays the r and nRMSE values achieved for the moments within both TD and CTEV groups using a DNN. The model predicted joint moments for TD children with an average correlation r of 0.75 and a nRMSE of 22.83 %. Similarly, for CTEV, the mean r was 0.76 and the nRMSE was 23.9 %.
Table 4.
r, nRMSE values of predicted continuous moments using deep neural network.
| Moment | TD |
CTEV |
|||
|---|---|---|---|---|---|
| r | nRMSE % | r | nRMSE % | ||
| Hip | flexion-extension | 0.82 | 13.51 | 0.91 | 12.94 |
| adduction-abduction | 0.79 | 27.00 | 0.81 | 21.12 | |
| Knee | flexion-extension | 0.73 | 23.24 | 0.83 | 25.32 |
| varus-valgus | 0.53 | 34.91 | 0.33 | 32.91 | |
| Ankle | dorsi-plantar flexion | 0.79 | 15.15 | 0.86 | 20.34 |
| inversion - eversion | 0.84 | 23.19 | 0.83 | 30.78 | |
Considering the evaluation parameters and the presence of artifacts random forest regressor and deep neural networks provided the most accurate predictions of joint moments for TD and CTEV children.
4. Discussion
The objective of this study was to develop and train an ML model for predicting joint moments based on joint kinematics as input. The findings reveal a higher accuracy in predicting sagittal joint moments compared to coronal joint moments. Additionally, the accuracy varied between TD and children with CTEV. Various ML models exhibited different levels of accuracy, with random forest regressor and deep neural networks emerging as the most effective predictors for both TD children and CTEV children.
Study9 employed one-dimensional convolutional neural networks to joint moments, using kinematics during gait in cerebral palsy patients and TD individuals. Joint moment predictions exhibited a nRMSE range of 18.02 %–13.58 % for the CP group and 12.55 %–8.58 % for the TD group. While indicating the potential for ML-based joint moment prediction to replace conventional calculation methods, especially for CP patients. The study also highlighted the current limitations in prediction accuracy, constraining its immediate clinical applicability. However, our study went beyond just predicting joint moments. We also compared seven different ML algorithms for this task. A similar study8 examined the performance of two neural network types, the feedforward network, and the recurrent long short-term memory network, in predicting lower limb joint moments during gait using joint angles from optical motion capture as input data. The average prediction accuracy across all parameters was higher with the FF network (r = 0.963) compared to the LSTM network (r = 0.935). For TD subjects, the nRMSE ranged from 12.14 % to 15.00 % for the FF neural network. Our findings indicate higher nRMSE values compared to prior studies. Specifically, the mean nRMSE errors ranged from 17.7 % to 43.7 % with polynomial regression, 14.13 %–30.53 % with random forest regression, 24.8 %–32.5 % with gradient boosting regression, and 13.5 % to 34.9 % with DNN regression for TD children. The error might be attributed to previous studies focusing on adults, whereas our research targets the pediatric population. Children tend to move at a faster pace,15 potentially leading to variations in GRF and consequently altering joint moments. Higher GRF was also observed in children compared to adults.16 Additionally, existing studies17 have demonstrated disparities in joint moments between adults and children. Furthermore, children show less gait consistency and it improves gradually with age, reaching stability around or after skeletal maturity.18 Similar error values were observed, with slightly higher values obtained for CTEV children due to variations in both kinematics and GRF among the CTEV group.
Similar study19 focused on collecting electromyography data from 10 muscles and angle data from 5 joints in the right lower limbs of healthy participants to analyze four joint moments. These data served as inputs for an ANN model, which was trained to predict the joint moments. The prediction accuracy was assessed using the nRMSE%, which was 7.89 %, and the correlation coefficient of 0.96 between the predicted joint moments and those obtained from multibody dynamics analysis. Additional study11 created an ANN model designed to predict external knee flexion moments and external knee adduction moments, utilizing data collected from wearable sensors. The effectiveness of these predictions was evaluated, showing a correlation coefficient of 0.71 ± 0.26 and a relative rRMSE of 22.3 ± 8.3 %. While these studies share similarities, they employed distinct methodologies.
The limitation of our study lies in the dataset used. Although sufficient, the machine learning algorithm should continually be developed further with data from an increasing number of cases. Another limitation is that only the kinematics of the lower body were considered as input. The machine learning model could be enhanced by incorporating full-body kinematics.
5. Conclusion
The findings of the study suggest that ML-based prediction of joint moments from kinematics holds promise as an alternative to conventional methods in gait analysis for both TD children and those with mobility challenges. However, current limitations in the accuracy of prediction errors hinder the immediate clinical application of ML-based techniques for decision-making in a pediatric population. The choice of the ML model proved crucial in predicting joint moments accurately. Exploring ML-based joint moment prediction using kinematics input for children facing mobility challenges represents an ambitious avenue for future research.
Funding/sponsorship
None.
Ethical approval
Approved.
Declaration of patient consent form
Provided.
Ethical approval and patient consent
The study has been conducted in accordance with the ethical principles mentioned in the Declaration of Helsinski (2013)
Funding/sponsorship
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Guardian/patient's consent
Informed consent was provided to the Guardian/Patient's consent.
CRediT authorship contribution statement
Rohan Kothurkar: Conceptualization, Software, Formal analysis, Writing – original draft. Mayuri Gad: Data collection, Resources. Abhiroop Padate: Formal analysis, Software. Chasanal Rathod: Writing – review & editing, Validation, Supervision. Atul Bhaskar: Writing – review & editing, Supervision. Ramesh Lekurwale: Validation, Supervision. John Rose: Validation, Supervision.
Declaration of competing interest
None.
Acknowledgment
None.
References
- 1.Felson D.T. Osteoarthritis as a disease of mechanics. Osteoarthr Cartil. 2013;21(1):10–15. doi: 10.1016/J.JOCA.2012.09.012. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Saxby D.J., Modenese L., Bryant A.L., et al. Tibiofemoral contact forces during walking, running and sidestepping. Gait Posture. 2016;49:78–85. doi: 10.1016/j.gaitpost.2016.06.014. [DOI] [PubMed] [Google Scholar]
- 3.Kothurkar R., Lekurwale R., Gad M., Rathod C.M. Estimation and Comparison of knee joint contact forces during heel contact and heel rise deep squatting. Indian J Orthop. 2022;57(2):310–318. doi: 10.1007/S43465-022-00798-Y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Kothurkar R., Lekurwale R., Gad M. Proceedings of International Conference on Intelligent Manufacturing and Automation. Springer; Singapore: 2023. Comparison of methods for predicting muscle activations and knee joint contact forces during squatting using OpenSim; pp. 533–540. [DOI] [Google Scholar]
- 5.Zeighami A., Dumas R., Aissaoui R. Knee loading in OA subjects is correlated to flexion and adduction moments and to contact point locations. Sci Rep. 2021;11(1):1–9. doi: 10.1038/s41598-021-87978-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Kutzner I., Trepczynski A., Heller M.O., Bergmann G. Knee adduction moment and medial contact force – facts about their correlation during gait. PLoS One. 2013;8(12) doi: 10.1371/JOURNAL.PONE.0081036. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Creaby M.W. It's not all about the knee adduction moment: the role of the knee flexion moment in medial knee joint loading. Osteoarthr Cartil. 2015;23(7):1038–1040. doi: 10.1016/j.joca.2015.03.032. [DOI] [PubMed] [Google Scholar]
- 8.Mundt M., Koeppe A., David S., Bamer F., Potthast W., Markert B. Prediction of ground reaction force and joint moments based on optical motion capture data during gait. Med Eng Phys. 2020;86:29–34. doi: 10.1016/j.medengphy.2020.10.001. [DOI] [PubMed] [Google Scholar]
- 9.Ozates M.E., Karabulut D., Salami F., Wolf S.I., Arslan Y.Z. Machine learning-based prediction of joint moments based on kinematics in patients with cerebral palsy. J Biomech. 2023;155(May) doi: 10.1016/j.jbiomech.2023.111668. [DOI] [PubMed] [Google Scholar]
- 10.Perrone M., Mell S.P., Martin J., Nho S.J., Malloy P. Machine learning-based prediction of hip joint moment in healthy subjects, patients and post-operative subjects. Comput Methods Biomech Biomed Engin. 2024;0(0):1–5. doi: 10.1080/10255842.2024.2310732. [DOI] [PubMed] [Google Scholar]
- 11.Stetter B.J., Krafft F.C., Ringhof S., Stein T., Sell S. A machine learning and wearable sensor based approach to estimate external knee flexion and adduction moments during various locomotion tasks. Front Bioeng Biotechnol. 2020;8 doi: 10.3389/FBIOE.2020.00009/BIBTEX. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.De Brabandere A., Emmerzaal J., Timmermans A., Jonkers I., Vanwanseele B., Davis J. A machine learning approach to estimate hip and knee joint loading using a mobile phone-embedded IMU. Front Bioeng Biotechnol. 2020;8 doi: 10.3389/FBIOE.2020.00320/BIBTEX. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Giarmatzis G., Zacharaki E.I., Moustakas K. Real-time prediction of joint forces by motion capture and machine learning. Sensors. 2020;20:6933. doi: 10.3390/S20236933. 2020;20(23):6933. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Pedregosa F., Varoquaux G., Gramfort A., et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12:2825–2830. [Google Scholar]
- 15.Aloba A., Luc A., Woodward J., et al. Quantifying differences between child and adult motion based on gait features. Lect Notes Comput Sci. 2019;11573(LNCS):385–402. doi: 10.1007/978-3-030-23563-5_31/FIGURES/2. [DOI] [Google Scholar]
- 16.Cowgill L.W., Warrener A., Pontzer H., Ocobock C. Waddling and toddling: the biomechanical effects of an immature gait. Am J Phys Anthropol. 2010;143(1):52–61. doi: 10.1002/AJPA.21289. [DOI] [PubMed] [Google Scholar]
- 17.Chester V.L., Wrigley A.T. The identification of age-related differences in kinetic gait parameters using principal component analysis. Clin Biomech. 2008;23(2):212–220. doi: 10.1016/J.CLINBIOMECH.2007.09.007. [DOI] [PubMed] [Google Scholar]
- 18.Sangeux M., Passmore E., Graham H.K., Tirosh O. The gait standard deviation, a single measure of kinematic variability. Gait Posture. 2016;46:194–200. doi: 10.1016/J.GAITPOST.2016.03.015. [DOI] [PubMed] [Google Scholar]
- 19.Xiong B., Zeng N., Li H., et al. Intelligent prediction of human lower extremity joint moment: an artificial neural network approach. IEEE Access. 2019;7:29973–29980. doi: 10.1109/ACCESS.2019.2900591. [DOI] [Google Scholar]