Prediction of ground reaction forces and moments during walking in children with cerebral palsy

Self-Supervised Learning Improves Accuracy and Data Efficiency for IMU-Based Ground Reaction Force Estimation

OBJECTIVE

Recent deep learning techniques hold promise to enable IMU-driven kinetic assessment; however, they require large extents of ground reaction force (GRF) data to serve as labels for supervised model training. We thus propose using existing self-supervised learning (SSL) techniques to leverage large IMU datasets to pre-train deep learning models, which can improve the accuracy and data efficiency of IMU-based GRF estimation.

METHODS

We performed SSL by masking a random portion of the input IMU data and training a transformer model to reconstruct the masked portion. We systematically compared a series of masking ratios across three pre-training datasets that included real IMU data, synthetic IMU data, or a combination of the two. Finally, we built models that used pre-training and labeled data to estimate GRF during three prediction tasks: overground walking, treadmill walking, and drop landing.

RESULTS

When using the same amount of labeled data, SSL pre-training significantly improved the accuracy of 3-axis GRF estimation during walking compared to baseline models trained by conventional supervised learning. Fine-tuning SSL model with 1-10% of walking data yielded comparable accuracy to training baseline model with 100% of walking data. The optimal masking ratio for SSL is 6.25-12.5%.

CONCLUSION

SSL leveraged large real and synthetic IMU datasets to increase the accuracy and data efficiency of deep-learning-based GRF estimation, reducing the need for labeled data.

SIGNIFICANCE

This work, with its open-source code and models, may unlock broader use cases of IMU-driven kinetic assessment by mitigating the scarcity of GRF measurements in practical applications.

Evaluation of ground reaction forces and centers of pressure predicted by AnyBody Modeling System during load reaching/handling activities and effects of the prediction errors on model-estimated spinal loads

Full-body and lower-extremity human musculoskeletal models require feet ground reaction forces (GRFs) and centers of pressure (CoPs) as inputs to predict muscle forces and joint loads. GRFs/CoPs are traditionally measured via floor-mounted forceplates that are usually restricted to research laboratories thus limiting their applicability in real occupational and clinical setups. Alternatively, GRFs/CoPs can be estimated via inverse dynamic approaches as also implemented in the Anybody Modeling System (AnyBody Technology, Aalborg, Denmark). The accuracy of Anybody in estimating GRFs/CoPs during load-handling/reaching activities and the effect of its prediction errors on model-estimated spinal loads remain to be investigated. Twelve normal- and over-weight individuals performed total of 480 static load-handling/reaching activities while measuring (by forceplates) and predicting (by AnyBody) their GRFs/CoPs. Moreover, the effects of GRF/CoP prediction errors on the estimated spinal loads were evaluated by inputting measured or predicted GRFs/CoPs into subject-specific musculoskeletal models. Regardless of the subject groups (normal-weight or overweight) and tasks (load-reaching or load-handling), results indicated great agreements between the measured and predicted GRFs (normalized root-mean-squared error, nRMSEs < 14% and R2 > 0.90) and between their model-estimated spinal loads (nRMSEs < 14% and R2 > 0.83). These agreements were good but relatively less satisfactory for CoPs (nRMSEs < 17% and 0.57 < R2 < 0.68). The only exception, requiring a more throughout investigation, was the situation when the ground-foot contact was significantly reduced during the activity. It appears that occupational/clinical investigations performed in real workstation/clinical setups with no access to forceplates may benefit from the AnyBody GRF/CoP prediction tools for a wide range of load-reaching/handling activities.

Self-Supervised Learning Improves Accuracy and Data Efficiency for IMU-Based Ground Reaction Force Estimation

Objective

Recent deep learning techniques hold promise to enable IMU-driven kinetic assessment; however, they require large extents of ground reaction force (GRF) data to serve as labels for supervised model training. We thus propose using existing self-supervised learning (SSL) techniques to leverage large IMU datasets to pre-train deep learning models, which can improve the accuracy and data efficiency of IMU-based GRF estimation.

Methods

We performed SSL by masking a random portion of the input IMU data and training a transformer model to reconstruct the masked portion. We systematically compared a series of masking ratios across three pre-training datasets that included real IMU data, synthetic IMU data, or a combination of the two. Finally, we built models that used pre-training and labeled data to estimate GRF during three prediction tasks: overground walking, treadmill walking, and drop landing.

Results

When using the same amount of labeled data, SSL pre-training significantly improved the accuracy of 3-axis GRF estimation during walking compared to baseline models trained by conventional supervised learning. Fine-tuning SSL model with 1-10% of walking data yielded comparable accuracy to training baseline model with 100% of walking data. The optimal masking ratio for SSL is 6.25-12.5%.

Conclusion

SSL leveraged large real and synthetic IMU datasets to increase the accuracy and data efficiency of deep-learning-based GRF estimation, reducing the need for labeled data.

Significance

This work, with its open-source code and models, may unlock broader use cases of IMU-driven kinetic assessment by mitigating the scarcity of GRF measurements in practical applications.

Ground Contact Force and Moment Estimation for Human-Exoskeleton Systems Using Dynamic Decoupled Coordinate System and Minimum Energy Hypothesis

Estimating the contact forces and moments (CFMs) between exoskeletons' feet and the ground is a prerequisite for calculating exoskeletons' joint moments. However, comfortable, portable, and high-precision force sensors for CFM detection are difficult to design and manufacture. In addition, there are many unknown CFM components (six force components and six moment components in the double-support phase). These reasons make it challenging to estimate CFMs precisely. In this paper, we propose a novel method for estimating these CFMs based on a proposed dynamic decoupled coordinate system (DDCS) and the minimum energy hypothesis. By decomposing these CFMs into a DDCS, the number of unknowns can be significantly reduced from twelve to two. Meanwhile, the minimum energy hypothesis provides a relatively reliable target for optimizing the remaining two unknown variables. We verify the accuracy of this method using a public data set about human walking. The validation shows that the proposed method is capable of estimating CFMs. This study provides a practical way to estimate the CFMs under the soles, which contributes to reducing the research and development costs of exoskeletons by avoiding the need for expensive plantar sensors. The sensor-free approach also reduces the dependence on high-precision, portable, and comfortable CFM detection sensors, which are usually difficult to design.

Machine learning-based prediction of joint moments based on kinematics in patients with cerebral palsy

Joint moments during gait provide valuable information for clinical decision-making in patients with cerebral palsy (CP). Joint moments are calculated based on ground reaction forces (GRF) using inverse dynamics models. Obtaining GRF from patients with CP is challenging. Typically developed (TD) individuals' joint moments were predicted from joint angles using machine learning, but no such study has been conducted on patients with CP. Accordingly, we aimed to predict the dorsi-plantar flexion, knee flexion-extension, hip flexion-extension, and hip adduction-abduction moments based on the trunk, pelvis, hip, knee, and ankle kinematics during gait in patients with CP and TD individuals using one-dimensional convolutional neural networks (CNN). The anonymized retrospective gait data of 329 TD (26 years ± 14, mass: 70 kg ± 15, height: 167 cm ± 89) and 917 CP (17 years ± 9, mass:47 kg ± 19, height:153 cm ± 36) individuals were evaluated and after applying inclusion-exclusion criteria, 132 TD and 622 CP patients with spastic diplegia were selected. We trained specific CNN models and evaluated their performance using isolated test subject groups based on normalized root mean square error (nRMSE) and Pearson correlation coefficient (PCC). Joint moments were predicted with nRMSE between 18.02% and 13.58% for the CP and between 12.55% and 8.58% for the TD groups, whereas with PCC between 0.85 and 0.93 for the CP and between 0.94 and 0.98 for the TD groups. Machine learning-based joint moment prediction from kinematics could replace conventional moment calculation in CP patients in the future, but the current level of prediction errors restricts its use for clinical decision-making today.