Dataset Information

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

ABSTRACT:

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.

SUBMITTER: Tan T

PROVIDER: S-EPMC10849467 | biostudies-literature | 2024 Jan

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

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

Tan Tian T Shull Peter B PB Hicks Jenifer L JL Uhlrich Scott D SD Chaudhari Akshay S AS

bioRxiv : the preprint server for biology 20240125

<h4>Objective</h4>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.<h4>Methods</h4>We performed SSL by masking a random por ...[more]

PMID: 38328126

Similar Datasets

Project description:Legged robots have become popular in recent years due to their ability to locomote on rough terrains; these robots are able to walk on narrow stepping-stones, go upstairs, and explore soft ground such as sand. Ground reaction force (GRF) is the force exerted on the body by the ground when they are in contact. This is a key element and is widely used for programming the locomotion of the legged robots. Being capable of estimating the GRF is advantageous over measuring it with the actual sensor system. Estimating allows one to simplify the system, and it is meant to be capable of prediction, and so on. In this article, we present a neural network approach for GRF estimation for the legged robot system. In order to fundamentally study the GRF estimation of the robot leg, we demonstrate our approach for a single-legged robot with a degree of freedom (DoF) of two with hip and knee joints on a flat-surface. The first joint is directly driven from the actuator, and another joint is belt-pulley driven from the second actuator to take advantage of the long range of motion. The neural network is designed to estimate GRF without attaching force sensors such as load cells, and the encoder is the only sensor used for the estimation. We propose a two-staged multi-layer perceptron (MLP) solution based on supervised learning to estimate GRF in the physical-world. The first stage of the MLP model is trained using datasets from the simulation, enabling it to estimate the simulation-staged GRF. The second stage of the MLP model is trained in the physical world using the simulation-staged GRF obtained from the first stage MLP as the input. This approach enables the second stage MLP to bridge the simulation to the physical world. The root mean squared error (RMSE) is 0.9949 N on the validation datasets in the best case. The performance of the trained network is evaluated when the robot follows trajectories that are not used in training the two-stage GRF estimation network.

Project description:PurposeThe association between lower-extremity loading and clinically relevant knee symptoms at different time points after anterior cruciate ligament reconstruction (ACLR) is unclear. Vertical ground reaction force (vGRF) from walking was compared between individuals with and without clinically relevant knee symptoms in three cohorts: <12 months post-ACLR, 12-24 months post-ACLR, and >24 months post-ACLR.MethodsOne hundred twenty-eight individuals with unilateral ACLR were classified as symptomatic or asymptomatic, based on previously defined cutoff values for the Knee Osteoarthritis and Injury Outcome Score (<12 months post-ACLR [symptomatic n = 28, asymptomatic n = 24]; 12-24 months post-ACLR [symptomatic n = 15, asymptomatic n = 15], and >24 months post-ACLR [symptomatic, n = 13; asymptomatic, n = 33]). Vertical ground reaction force exerted on the ACLR limb was collected during walking gait, and functional analyses of variance were used to evaluate the effects of symptoms and time post-ACLR on vGRF throughout stance phase (α = 0.05).ResultsSymptomatic individuals, <12 months post-ACLR, demonstrated less vGRF during both vGRF peaks (i.e., weight acceptance and propulsion) and greater vGRF during midstance, compared to asymptomatic individuals. Vertical ground reaction force characteristics were not different between symptomatic and asymptomatic individuals for most of stance in individuals between 12 and 24 months post-ACLR. Symptomatic individuals who were >24 months post-ACLR, exhibited greater vGRF during both peaks, but lesser vGRF during midstance, compared to asymptomatic individuals.ConclusionRelative to asymptomatic individuals, symptomatic individuals are more likely to underload the ACLR limb early after ACLR (i.e., <12 months) during both vGRF peaks, but overload the ACLR limb, during both vGRF peaks, at later time points (i.e., >24 months). We propose these differences in lower-extremity loading during walking might have implications for long-term knee health, and should be considered when designing therapeutic interventions for individuals with an ACLR.

Project description:With a spatial resolution of tens of microns, ultrasound localization microscopy (ULM) reconstructs microvascular structures and measures intravascular flows by tracking microbubbles (1-5 μm) in contrast enhanced ultrasound (CEUS) images. Since the size of CEUS bubble traces, e.g. 0.5-1 mm for ultrasound with a wavelength λ = 280 μm, is typically two orders of magnitude larger than the bubble diameter, accurately localizing microbubbles in noisy CEUS data is vital to the fidelity of the ULM results. In this paper, we introduce a residual learning based supervised super-resolution blind deconvolution network (SupBD-net), and a new loss function for a self-supervised blind deconvolution network (SelfBD-net), for detecting bubble centers at a spatial resolution finer than λ/10. Our ultimate purpose is to improve the ability to distinguish closely located microvessels and the accuracy of the velocity profile measurements in macrovessels. Using realistic synthetic data, the performance of these methods is calibrated and compared against several recently introduced deep learning and blind deconvolution techniques. For bubble detection, errors in bubble center location increase with the trace size, noise level, and bubble concentration. For all cases, SupBD-net yields the least error, keeping it below 0.1 λ. For unknown bubble trace morphology, where all the supervised learning methods fail, SelfBD-net can still maintain an error of less than 0.15 λ. SupBD-net also outperforms the other methods in separating closely located bubbles and parallel microvessels. In macrovessels, SupBD-net maintains the least errors in the vessel radius and velocity profile after introducing a procedure that corrects for terminated tracks caused by overlapping traces. Application of these methods is demonstrated by mapping the cerebral microvasculature of a neonatal pig, where neighboring microvessels separated by 0.15 λ can be readily distinguished by SupBD-net and SelfBD-net, but not by the other techniques. Hence, the newly proposed residual learning based methods improve the spatial resolution and accuracy of ULM in micro- and macro-vessels.

Dataset Information

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

Objective

Methods

Results

Conclusion

Significance

Publications

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

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets