Estimation of vertical ground reaction force during running using neural network model and uniaxial accelerometer

Biomechanical Effects of Manipulating Preferred Cadence During Treadmill Walking in Patients With ACL Reconstruction

BACKGROUND

Abnormal gait is common after anterior cruciate ligament reconstruction (ACLR) which may influence osteoarthritis risk in this population. Yet few gait retraining options currently exist in ACLR rehabilitation. Cueing cadence changes is a simple, low-cost method that can alter walking mechanics in healthy adults, but few studies have tested its effectiveness in an ACLR population. Here, we evaluated the acute effects of altering cadence on knee mechanics in patients 9 to 12 months post ACLR.

HYPOTHESIS

Cueing larger steps will facilitate larger knee angles and moments, while cueing smaller steps would induce smaller knee angles and moments.

STUDY DESIGN

Randomized cross-sectional design.

LEVEL OF EVIDENCE

Level 3.

METHODS

Twenty-eight patients with unilateral ACLR underwent gait assessments on a treadmill at preferred pace. Preferred walking gait was assessed first to obtain preferred cadence. Participants then completed trials while matching an audible beat set to 90% and 110% of preferred cadence in a randomized order. Three-dimensional sagittal and frontal plane biomechanics were evaluated bilaterally.

RESULTS

Compared with preferred cadence, cueing larger steps induced larger peak knee flexion moments (KFMs) and knee extension excursions bilaterally (P < 0.01), whereas cueing smaller steps only reduced knee flexion excursions (P < 0.01). Knee adduction moments remain unchanged across conditions and were similar between limbs (P > 0.05). Peak KFMs and excursions were smaller in the injured compared with uninjured limb (P < 0.01).

CONCLUSION

Frontal plane gait outcomes were unchanged across conditions suggesting acute cadence manipulations result in mainly sagittal plane adaptations. Follow-up studies using a longitudinal cadence biofeedback paradigm may be warranted to elucidate the utility of this gait retraining strategy after ACLR.

CLINICAL RELEVANCE

Cueing changes in walking cadence can target sagittal plane knee loading and joint range of motion in ACLR participants. This strategy may offer high clinical translatability given it requires relatively minimal equipment (ie, free metronome app) outside of a treadmill.

Rethinking running biomechanics: a critical review of ground reaction forces, tibial bone loading, and the role of wearable sensors

This study presents a comprehensive review of the correlation between tibial acceleration (TA), ground reaction forces (GRF), and tibial bone loading, emphasizing the critical role of wearable sensor technology in accurately measuring these biomechanical forces in the context of running. This systematic review and meta-analysis searched various electronic databases (PubMed, SPORTDiscus, Scopus, IEEE Xplore, and ScienceDirect) to identify relevant studies. It critically evaluates existing research on GRF and tibial acceleration (TA) as indicators of running-related injuries, revealing mixed findings. Intriguingly, recent empirical data indicate only a marginal link between GRF, TA, and tibial bone stress, thus challenging the conventional understanding in this field. The study also highlights the limitations of current biomechanical models and methodologies, proposing a paradigm shift towards more holistic and integrated approaches. The study underscores wearable sensors' potential, enhanced by machine learning, in transforming the monitoring, prevention, and rehabilitation of running-related injuries.

Estimation of Kinetics Using IMUs to Monitor and Aid in Clinical Decision-Making during ACL Rehabilitation: A Systematic Review

After an ACL injury, rehabilitation consists of multiple phases, and progress between these phases is guided by subjective visual assessments of activities such as running, hopping, jump landing, etc. Estimation of objective kinetic measures like knee joint moments and GRF during assessment can help physiotherapists gain insights on knee loading and tailor rehabilitation protocols. Conventional methods deployed to estimate kinetics require complex, expensive systems and are limited to laboratory settings. Alternatively, multiple algorithms have been proposed in the literature to estimate kinetics from kinematics measured using only IMUs. However, the knowledge about their accuracy and generalizability for patient populations is still limited. Therefore, this article aims to identify the available algorithms for the estimation of kinetic parameters using kinematics measured only from IMUs and to evaluate their applicability in ACL rehabilitation through a comprehensive systematic review. The papers identified through the search were categorized based on the modelling techniques and kinetic parameters of interest, and subsequently compared based on the accuracies achieved and applicability for ACL patients during rehabilitation. IMUs have exhibited potential in estimating kinetic parameters with good accuracy, particularly for sagittal movements in healthy cohorts. However, several shortcomings were identified and future directions for improvement have been proposed, including extension of proposed algorithms to accommodate multiplanar movements and validation of the proposed techniques in diverse patient populations and in particular the ACL population.

Integrating an LSTM framework for predicting ankle joint biomechanics during gait using inertial sensors

The ankle joint plays a crucial role in gait, facilitating the articulation of the lower limb, maintaining foot-ground contact, balancing the body, and transmitting the center of gravity. This study aimed to implement long short-term memory (LSTM) networks for predicting ankle joint angles, torques, and contact forces using inertial measurement unit (IMU) sensors. Twenty-five healthy participants were recruited. Two IMU sensors were attached to the foot dorsum and the vertical axis of the distal anteromedial tibia in the right lower limb to record acceleration and angular velocity during running. We proposed a LSTM-MLP (multilayer perceptron) model for training time-series data from IMU sensors and predicting ankle joint biomechanics. The model underwent validation and testing using a custom nested k-fold cross-validation process. The average values of the coefficient of determination (R2), mean absolute error (MAE), and mean squared error (MSE) for ankle dorsiflexion joint and moment, subtalar inversion joint and moment, and ankle joint contact forces were 0.89 ± 0.04, 0.75 ± 1.04, and 2.96 ± 4.96 for walking, and 0.87 ± 0.07, 0.88 ± 1.26, and 4.1 ± 7.17 for running, respectively. This study demonstrates that IMU sensors, combined with LSTM neural networks, are invaluable tools for evaluating ankle joint biomechanics in lower limb pathological diagnosis and rehabilitation, offering a cost-effective and versatile alternative to traditional experimental settings.

Acceleration-Based Estimation of Vertical Ground Reaction Forces during Running: A Comparison of Methods across Running Speeds, Surfaces, and Foot Strike Patterns

Twenty-seven methods of estimating vertical ground reaction force first peak, loading rate, second peak, average, and/or time series from a single wearable accelerometer worn on the shank or approximate center of mass during running were compared. Force estimation errors were quantified for 74 participants across different running surfaces, speeds, and foot strike angles and biases, repeatability coefficients, and limits of agreement were modeled with linear mixed effects to quantify the accuracy, reliability, and precision. Several methods accurately and reliably estimated the first peak and loading rate, however, none could do so precisely (the limits of agreement exceeded ±65% of target values). Thus, we do not recommend first peak or loading rate estimation from accelerometers with the methods currently available. In contrast, the second peak, average, and time series could all be estimated accurately, reliably, and precisely with several different methods. Of these, we recommend the 'Pogson' methods due to their accuracy, reliability, and precision as well as their stability across surfaces, speeds, and foot strike angles.

Foot Pronation Prediction with Inertial Sensors during Running: A Preliminary Application of Data-Driven Approaches

Abnormal foot postures may affect foot movement and joint loading during locomotion. Investigating foot posture alternation during running could contribute to injury prevention and foot mechanism study. This study aimed to develop feature-based and deep learning algorithms to predict foot pronation during prolonged running. Thirty-two recreational runners have been recruited for this study. Nine-axial inertial sensors were attached to the right dorsum of the foot and the vertical axis of the distal anteromedial tibia. This study employed feature-based machine learning algorithms, including support vector machine (SVM), extreme gradient boosting (XGBoost), random forest, and deep learning, i.e., one-dimensional convolutional neural networks (CNN1D), to predict foot pronation. A custom nested k-fold cross-validation was designed for hyper-parameter tuning and validating the model's performance. The XGBoot classifier achieved the best accuracy using acceleration and angular velocity data from the foot dorsum as input. Accuracy and the area under curve (AUC) were 74.7 ± 5.2% and 0.82 ± 0.07 for the subject-independent model and 98 ± 0.4% and 0.99 ± 0 for the record-wise method. The test accuracy of the CNN1D model with sensor data at the foot dorsum was 74 ± 3.8% for the subject-wise approach with an AUC of 0.8 ± 0.05. This study found that these algorithms, specifically for the CNN1D and XGBoost model with inertial sensor data collected from the foot dorsum, could be implemented into wearable devices, such as a smartwatch, for monitoring a runner's foot pronation during long-distance running. It has the potential for running shoe matching and reducing or preventing foot posture-induced injuries.

Predicting vertical ground reaction force in rearfoot running: A wavelet neural network model and factor loading

This study proposed a simple method for selecting input variables by factor loading and inputting these variables into a wavelet neural network (WNN) model to predict vertical ground reaction force (vGRF). The kinematic data and vGRF of 9 rearfoot strikers at 12, 14, and 16 km/h were collected using a motion capture system and an instrumented treadmill. The input variables were screened by factor loading and utilized to predict vGRF with the WNN. Nine kinematic variables were selected, corresponding to nine principal components, mainly focusing on the knee and ankle joints. The prediction results of vGRF were effective and accurate at different speeds, namely, the coefficient of multiple correlation (CMC) > 0.98 (0.984-0.988), the normalized root means square error (NRMSE) < 15% (9.34-11.51%). The NRMSEs of impact force (8.18-10.01%), active force (4.92-7.42%), and peak time (7.16-12.52%) were less than 15%. There was a small number (peak, 4.12-6.18%; time, 4.71-6.76%) exceeding the 95% confidence interval (CI) using the Bland-Altman method. The knee joint was the optimal location for estimating vGRF, followed by the ankle. There were high accuracy and agreement for predicting vGRF with the peak and peak time at 12, 14, and 16 km/h. Therefore, factor loading could be a valid method to screen kinematic variables in artificial neural networks.

Limiting the Use of Electromyography and Ground Reaction Force Data Changes the Magnitude and Ranking of Modelled Anterior Cruciate Ligament Forces

Neuromusculoskeletal models often require three-dimensional (3D) body motions, ground reaction forces (GRF), and electromyography (EMG) as input data. Acquiring these data in real-world settings is challenging, with barriers such as the cost of instruments, setup time, and operator skills to correctly acquire and interpret data. This study investigated the consequences of limiting EMG and GRF data on modelled anterior cruciate ligament (ACL) forces during a drop–land–jump task in late-/post-pubertal females. We compared ACL forces generated by a reference model (i.e., EMG-informed neural mode combined with 3D GRF) to those generated by an EMG-informed with only vertical GRF, static optimisation with 3D GRF, and static optimisation with only vertical GRF. Results indicated ACL force magnitude during landing (when ACL injury typically occurs) was significantly overestimated if only vertical GRF were used for either EMG-informed or static optimisation neural modes. If 3D GRF were used in combination with static optimisation, ACL force was marginally overestimated compared to the reference model. None of the alternative models maintained rank order of ACL loading magnitudes generated by the reference model. Finally, we observed substantial variability across the study sample in response to limiting EMG and GRF data, indicating need for methods incorporating subject-specific measures of muscle activation patterns and external loading when modelling ACL loading during dynamic motor tasks.

Thermal Calibration of Triaxial Accelerometer for Tilt Measurement

The application of MEMS accelerometers used to measure inclination is constrained by their temperature dependence, and each accelerometer needs to be calibrated individually to increase stability and accuracy. This paper presents a calibration and thermal compensation method for triaxial accelerometers that aims to minimize cost and processing time while maintaining high accuracy. First, the number of positions to perform the calibration procedure is optimized based on the Levenberg-Marquardt algorithm, and then, based on this optimized calibration number, thermal compensation is performed based on the least squares method, which is necessary for environments with large temperature variations, since calibration parameters change at different temperatures. The calibration procedures and algorithms were experimentally validated on marketed accelerometers. Based on the optimized calibration method, the calibrated results achieved nearly 100 times improvement. Thermal drift calibration experiments on the triaxial accelerometer show that the thermal compensation scheme in this paper can effectively reduce drift in the temperature range of -40 °C to 60 °C. The temperature drifts of x- and y-axes are reduced from -13.2 and 11.8 mg to -0.9 and -1.1 mg, respectively. The z-axis temperature drift is reduced from -17.9 to 1.8 mg. We have conducted various experiments on the proposed calibration method and demonstrated its capacity to calibrate the sensor frame error model (SFEM) parameters. This research proposes a new low-cost and efficient strategy for increasing the practical applicability of triaxial accelerometers.

Inertial Sensor Location for Ground Reaction Force and Gait Event Detection Using Reservoir Computing in Gait

Inertial measurement units (IMUs) have shown promising outcomes for estimating gait event detection (GED) and ground reaction force (GRF). This study aims to determine the best sensor location for GED and GRF prediction in gait using data from IMUs for healthy and medial knee osteoarthritis (MKOA) individuals. In this study, 27 healthy and 18 MKOA individuals participated. Participants walked at different speeds on an instrumented treadmill. Five synchronized IMUs (Physilog®, 200 Hz) were placed on the lower limb (top of the shoe, heel, above medial malleolus, middle and front of tibia, and on medial of shank close to knee joint). To predict GRF and GED, an artificial neural network known as reservoir computing was trained using combinations of acceleration signals retrieved from each IMU. For GRF prediction, the best sensor location was top of the shoe for 72.2% and 41.7% of individuals in the healthy and MKOA populations, respectively, based on the minimum value of the mean absolute error (MAE). For GED, the minimum MAE value for both groups was for middle and front of tibia, then top of the shoe. This study demonstrates that top of the shoe is the best sensor location for GED and GRF prediction.

Estimation of the knee joint load using plantar pressure data measured by smart socks: A feasibility study

BACKGROUND

Unsupervised sports activities could cause traumas, about 70% of them are those of the low extremities. To avoid traumas, the athlete should be aware of dangerous forces acting within low extremity joints. Research in gait analysis indicated that plantar pressure alteration rate correlates with the gait pace. Thus, the changes in plantar pressure should correlate with the accelerations of extremities, and with the forces, acting in the joints. Smart socks provide a budget solution for the measurement of plantar pressure.

OBJECTIVE

To estimate the correlation between the plantar pressure, measured using smart socks, and forces, acting in the joints of the lower extremities.

METHODS

The research is case study based. The volunteer performed a set of squats. The arbitrary plantar pressure-related data were obtained using originally developed smart socks with embedded knitted pressure sensors. Simultaneously, the lower extremity motion data were recorded using two inertial measurement units, attached to the tight and the ankle, from which the forces acted in the knee joint were estimated. The simplest possible model of knee joint mechanics was used to estimate force.

RESULTS

The estimates of the plantar pressure and knee joint forces demonstrate a strong correlation (r= 0.75, P< 0.001). The established linear regression equation enables the calculation of the knee joint force with an uncertainty of 22% using the plantar pressure estimate. The accuracy of the classification of the joint force as excessive, i.e., being more than 90% of the maximal force, was 82%.

CONCLUSION

The results demonstrate the feasibility of the smart socks for the estimation of the forces in the knee joints. Smart socks therefore could be used to develop excessive joint force alert devices, that could replace less convenient inertial sensors.

Estimating Ground Reaction Forces from Two-Dimensional Pose Data: A Biomechanics-Based Comparison of AlphaPose, BlazePose, and OpenPose

The adoption of computer vision pose estimation approaches, used to identify keypoint locations which are intended to reflect the necessary anatomical landmarks relied upon by biomechanists for musculoskeletal modelling, has gained increasing traction in recent years. This uptake has been further accelerated by keypoint use as inputs into machine learning models used to estimate biomechanical parameters such as ground reaction forces (GRFs) in the absence of instrumentation required for direct measurement. This study first aimed to investigate the keypoint detection rate of three open-source pose estimation models (AlphaPose, BlazePose, and OpenPose) across varying movements, camera views, and trial lengths. Second, this study aimed to assess the suitability and interchangeability of keypoints detected by each pose estimation model when used as inputs into machine learning models for the estimation of GRFs. The keypoint detection rate of BlazePose was distinctly lower than that of AlphaPose and OpenPose. All pose estimation models achieved a high keypoint detection rate at the centre of an image frame and a lower detection rate in the true sagittal plane camera field of view, compared with slightly anteriorly or posteriorly located quasi-sagittal plane camera views. The three-dimensional ground reaction force, instantaneous loading rate, and peak force for running could be estimated using the keypoints of all three pose estimation models. However, only AlphaPose and OpenPose keypoints could be used interchangeably with a machine learning model trained to estimate GRFs based on AlphaPose keypoints resulting in a high estimation accuracy when OpenPose keypoints were used as inputs and vice versa. The findings of this study highlight the need for further evaluation of computer vision-based pose estimation models for application in biomechanical human modelling, and the limitations of machine learning-based GRF estimation models that rely on 2D keypoints. This is of particular relevance given that machine learning models informing athlete monitoring guidelines are being developed for application related to athlete well-being.

Predicting vertical ground reaction forces from 3D accelerometry using reservoir computers leads to accurate gait event detection

Accelerometers are low-cost measurement devices that can readily be used outside the lab. However, determining isolated gait events from accelerometer signals, especially foot-off events during running, is an open problem. We outline a two-step approach where machine learning serves to predict vertical ground reaction forces from accelerometer signals, followed by force-based event detection. We collected shank accelerometer signals and ground reaction forces from 21 adults during comfortable walking and running on an instrumented treadmill. We trained one common reservoir computer using segmented data using both walking and running data. Despite being trained on just a small number of strides, this reservoir computer predicted vertical ground reaction forces in continuous gait with high quality. The subsequent foot contact and foot off event detection proved highly accurate when compared to the gold standard based on co-registered ground reaction forces. Our proof-of-concept illustrates the capacity of combining accelerometry with machine learning for detecting isolated gait events irrespective of mode of locomotion.

Mechanical loading prediction through accelerometry data during walking and running

Currently, there is no way to assess mechanical loading variables such as peak ground reaction forces (pGRF) and peak loading rate (pLR) in clinical settings. The purpose of this study was to develop accelerometry-based equations to predict both pGRF and pLR during walking and running. One hundred and thirty one subjects (79 females; 76.9 ± 19.6kg) walked and ran at different speeds (2-14km·h^-1) on a force plate-instrumented treadmill while wearing accelerometers at their ankle, lower back and hip. Regression equations were developed to predict pGRF and pLR from accelerometry data. Leave-one-out cross-validation was used to calculate prediction accuracy and Bland-Altman plots. Our pGRF prediction equation was compared with a reference equation previously published. Body mass and peak acceleration were included for pGRF prediction and body mass and peak acceleration rate for pLR prediction. All pGRF equation coefficients of determination were above 0.96, and a good agreement between actual and predicted pGRF was observed, with a mean absolute percent error (MAPE) below 7.3%. Accuracy indices from our equations were better than previously developed equations. All pLR prediction equations presented a lower accuracy compared to those developed to predict pGRF. Walking and running pGRF can be predicted with high accuracy by accelerometry-based equations, representing an easy way to determine mechanical loading in free-living conditions. The pLR prediction equations yielded a somewhat lower prediction accuracy compared with the pGRF equations.

Recent Machine Learning Progress in Lower Limb Running Biomechanics With Wearable Technology: A Systematic Review

With the emergence of wearable technology and machine learning approaches, gait monitoring in real-time is attracting interest from the sports biomechanics community. This study presents a systematic review of machine learning approaches in running biomechanics using wearable sensors. Electronic databases were retrieved in PubMed, Web of Science, SPORTDiscus, Scopus, IEEE Xplore, and ScienceDirect. A total of 4,068 articles were identified via electronic databases. Twenty-four articles that met the eligibility criteria after article screening were included in this systematic review. The range of quality scores of the included studies is from 0.78 to 1.00, with 40% of articles recruiting participant numbers between 20 and 50. The number of inertial measurement unit (IMU) placed on the lower limbs varied from 1 to 5, mainly in the pelvis, thigh, distal tibia, and foot. Deep learning algorithms occupied 57% of total machine learning approaches. Convolutional neural networks (CNN) were the most frequently used deep learning algorithm. However, the validation process for machine learning models was lacking in some studies and should be given more attention in future research. The deep learning model combining multiple CNN and recurrent neural networks (RNN) was observed to extract different running features from the wearable sensors and presents a growing trend in running biomechanics.

Estimating Running Ground Reaction Forces from Plantar Pressure during Graded Running

Ground reaction forces (GRFs) describe how runners interact with their surroundings and provide the basis for computing inverse dynamics. Wearable technology can predict time−continuous GRFs during walking and running; however, the majority of GRF predictions examine level ground locomotion. The purpose of this manuscript was to predict vertical and anterior–posterior GRFs across different speeds and slopes. Eighteen recreationally active subjects ran on an instrumented treadmill while we collected GRFs and plantar pressure. Subjects ran on level ground at 2.6, 3.0, 3.4, and 3.8 m/s, six degrees inclined at 2.6, 2.8, and 3.0 m/s, and six degrees declined at 2.6, 2.8, 3.0, and 3.4 m/s. We estimated GRFs using a set of linear models and a recurrent neural network, which used speed, slope, and plantar pressure as inputs. We also tested eliminating speed and slope as inputs. The recurrent neural network outperformed the linear model across all conditions, especially with the prediction of anterior–posterior GRFs. Eliminating speed and slope as model inputs had little effect on performance. We also demonstrate that subject−specific model training can reduce errors from 8% to 3%. With such low errors, researchers can use these wearable−based GRFs to understand running performance or injuries in real−world settings.

Use of Wearable Technology to Measure Activity in Orthopaedic Trauma Patients: A Systematic Review

BACKGROUND

Patient-Reported Outcome Measures (PROMs) are widely used for measurement of functional outcomes after orthopaedic trauma. However, PROMs rely on patient collaboration and suffer from various types of bias. Wearable Activity Monitors (WAMs) are increasingly used to objectify functional assessment. The objectives of this systematic review were to identify and characterise the WAMs technology and metrics currently used for orthopaedic trauma research.

METHODS

PubMed and Embase biomedical literature search engines were queried. Eligibility criteria included: Human clinical studies published in the English language between 2010 and 2019 involving fracture management and WAMs. Variables collected from each article included: Technology used, vendor/product, WAM body location, metrics measured, measurement time period, year of publication, study geographic location, phase of treatment studied, fractures studied, number of patients studied, sex and age of the study subjects, and study level of evidence. Six investigators reviewed the resulting papers. Descriptive statistics of variables of interest were used to analyse the data.

RESULTS

One hundred and thirty-six papers were available for analysis, showing an increasing trend of publications per year. Accelerometry followed by plantar pressure insoles were the most commonly employed technologies. The most common location for WAM placement was insoles, followed by the waist. The most commonly studied fracture type was hip fractures followed by fragility fractures in general, ankle, "lower extremity", and tibial fractures. The rehabilitation phase following surgery was the most commonly studied period. Sleep duration, activity time or step counts were the most commonly reported WAM metrics. A preferred, clinically validated WAM metric was not identified.

CONCLUSIONS

WAMs have an increasing presence in the orthopaedic trauma literature. The optimal implementation of this technology and its use to understand patients' pre-injury and post-injury functions is currently insufficiently explored and represents an area that will benefit from future study.

SYSTEMATIC REVIEW REGISTRATION NUMBER

PROSPERO ID:210344.

SUPPLEMENTARY INFORMATION

The online version contains supplementary material available at 10.1007/s43465-022-00629-0.

Performance of machine learning models in estimation of ground reaction forces during balance exergaming

Background

Balance training exercise games (exergames) are a promising tool for reducing fall risk in elderly. Exergames can be used for in-home guided exercise, which greatly increases availability and facilitates independence. Providing biofeedback on weight-shifting during in-home balance exercise improves exercise efficiency, but suitable equipment for measuring weight-shifting is lacking. Exergames often use kinematic data as input for game control. Being able to useg such data to estimate weight-shifting would be a great advantage. Machine learning (ML) models have been shown to perform well in weight-shifting estimation in other settings. Therefore, the aim of this study was to investigate the performance of ML models in estimation of weight-shifting during exergaming using kinematic data.

Methods

Twelve healthy older adults (mean age 72 (± 4.2), 10 F) played a custom exergame that required repeated weight-shifts. Full-body 3D motion capture (3DMoCap) data and standard 2D digital video (2D-DV) was recorded. Weight shifting was directly measured by 3D ground reaction forces (GRF) from force plates, and estimated using a linear regression model, a long-short term memory (LSTM) model and a decision tree model (XGBoost). Performance was evaluated using coefficient of determination (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}R2) and root mean square error (RMSE).

Results

Results from estimation of GRF components using 3DMoCap data show a mean (± 1SD) RMSE (% total body weight, BW) of the vertical GRF component (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_z$$\end{document}Fz) of 4.3 (2.5), 11.1 (4.5), and 11.0 (4.7) for LSTM, XGBoost and LinReg, respectively. Using 2D-DV data, LSTM and XGBoost achieve mean RMSE (± 1SD) in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_z$$\end{document}Fz estimation of 10.7 (9.0) %BW and 19.8 (6.4) %BW, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}R2 was \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$>.97$$\end{document}>.97 for the LSTM in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_z$$\end{document}Fz component using 3DMoCap data, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$>.77$$\end{document}>.77 using 2D-DV data. For XGBoost, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_z$$\end{document}Fz\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}R2 was \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$>.86$$\end{document}>.86 using 3DMoCap data, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$>.56$$\end{document}>.56 using 2D-DV data.

Conclusion

This study demonstrates that an LSTM model can estimate 3-dimensional GRF components using 2D kinematic data extracted from standard 2D digital video cameras. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_z$$\end{document}Fz component is estimated more accurately than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_y$$\end{document}Fy and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_x$$\end{document}Fx components, especially when using 2D-DV data. Weight-shifting performance during exergaming can thus be extracted using kinematic data only, which can enable effective independent in-home balance exergaming.

The Use of Wearable Sensor Technology to Detect Shock Impacts in Sports and Occupational Settings: A Scoping Review

Shock impacts during activity may cause damage to the joints, muscles, bones, or inner organs. To define thresholds for tolerable impacts, there is a need for methods that can accurately monitor shock impacts in real-life settings. Therefore, the main aim of this scoping review was to present an overview of existing methods for assessments of shock impacts using wearable sensor technology within two domains: sports and occupational settings. Online databases were used to identify papers published in 2010-2020, from which we selected 34 papers that used wearable sensor technology to measure shock impacts. No studies were found on occupational settings. For the sports domain, accelerometry was the dominant type of wearable sensor technology utilized, interpreting peak acceleration as a proxy for impact. Of the included studies, 28 assessed foot strike in running, head impacts in invasion and team sports, or different forms of jump landings or plyometric movements. The included studies revealed a lack of consensus regarding sensor placement and interpretation of the results. Furthermore, the identified high proportion of validation studies support previous concerns that wearable sensors at present are inadequate as a stand-alone method for valid and accurate data on shock impacts in the field.

A Wearable System for the Estimation of Performance-Related Metrics during Running and Jumping Tasks

Athletic performance, technique assessment, and injury prevention are all important aspects in sports for both professional and amateur athletes. Wearable technology is attracting the research community’s interest because of its capability to provide real-time biofeedback to coaches and athletes when on the field and outside of more restrictive laboratory conditions. In this paper, a novel wearable motion sensor-based system has been designed and developed for athletic performance assessment during running and jumping tasks. The system consists of a number of components involving embedded systems (hardware and software), back-end analytics, information and communications technology (ICT) platforms, and a graphical user interface for data visualization by the coach. The system is able to provide automatic activity recognition, estimation of running and jumping metrics, as well as vertical ground reaction force (GRF) predictions, with sufficient accuracy to provide valuable information as regards training outcomes. The developed system is low-power, sufficiently small for real-world scenarios, easy to use, and achieves the specified communication range. The system’s high sampling rate, levels of accuracy and performance enables it as a performance evaluation tool able to support coaches and athletes in their real-world practice.

Joint angle estimation with wavelet neural networks

This paper presents a wavelet neural network (WNN) based method to reduce reliance on wearable kinematic sensors in gait analysis. Wearable kinematic sensors hinder real-time outdoor gait monitoring applications due to drawbacks caused by multiple sensor placements and sensor offset errors. The proposed WNN method uses vertical Ground Reaction Forces (vGRFs) measured from foot kinetic sensors as inputs to estimate ankle, knee, and hip joint angles. Salient vGRF inputs are extracted from primary gait event intervals. These selected gait inputs facilitate future integration with smart insoles for real-time outdoor gait studies. The proposed concept potentially reduces the number of body-mounted kinematics sensors used in gait analysis applications, hence leading to a simplified sensor placement and control circuitry without deteriorating the overall performance.

The Effect of Footwear, Running Speed, and Location on the Validity of Two Commercially Available Inertial Measurement Units During Running

Introduction: Most running-related injuries are believed to be caused by abrupt changes in training load, compounded by biomechanical movement patterns. Wearable technology has made it possible for runners to quantify biomechanical loads (e.g., peak positive acceleration; PPA) using commercially available inertial measurement units (IMUs). However, few devices have established criterion validity. The aim of this study was to assess the validity of two commercially available IMUs during running. Secondary aims were to determine the effect of footwear, running speed, and IMU location on PPA.

Materials and Methods: Healthy runners underwent a biomechanical running analysis on an instrumented treadmill. Participants ran at their preferred speed in three footwear conditions (neutral, minimalist, and maximalist), and at three speeds (preferred, +10%, −10%) in the neutral running shoes. Four IMUs were affixed at the distal tibia (IMeasureU-Tibia), shoelaces (RunScribe and IMeasureU-Shoe), and insole (Plantiga) of the right shoe. Pearson correlations were calculated for average vertical loading rate (AVLR) and PPA at each IMU location.

Results: The AVLR had a high positive association with PPA (IMeasureU-Tibia) in the neutral and maximalist (r = 0.70–0.72; p ≤ 0.001) shoes and in all running speed conditions (r = 0.71–0.83; p ≤ 0.001), but low positive association in the minimalist (r = 0.47; p < 0.05) footwear condition. Conversely, the relationship between AVLR and PPA (Plantiga) was high in the minimalist (r = 0.75; p ≤ 0.001) condition and moderate in the neutral (r = 0.50; p < 0.05) and maximalist (r = 0.57; p < 0.01) footwear. The RunScribe metrics demonstrated low to moderate positive associations (r = 0.40–0.62; p < 0.05) with AVLR across most footwear and speed conditions.

Discussion: Our findings indicate that the commercially available Plantiga IMU is comparable to a tibia-mounted IMU when acting as a surrogate for AVLR. However, these results vary between different levels of footwear and running speeds. The shoe-mounted RunScribe IMU exhibited slightly lower positive associations with AVLR. In general, the relationship with AVLR improved for the RunScribe sensor at slower speeds and improved for the Plantiga and tibia-mounted IMeasureU sensors at faster speeds.

Does Site Matter? Impact of Inertial Measurement Unit Placement on the Validity and Reliability of Stride Variables During Running: A Systematic Review and Meta-analysis

BACKGROUND

Inertial measurement units (IMUs) are used for running gait analysis in a variety of sports. These sensors have been attached at various locations to capture stride data. However, it is unclear if different placement sites affect the derived outcome measures.

OBJECTIVE

The aim of this systematic review and meta-analysis was to investigate the impact of placement on the validity and reliability of IMU-derived measures of running gait.

METHODS

Online databases SPORTDiscus with Full Text, CINAHL Complete, MEDLINE (EBSCOhost), EMBASE (Ovid) and Scopus were searched from the earliest record to 6 August 2020. Articles were included if they (1) used an IMU during running (2) reported spatiotemporal variables, peak ground reaction force (GRF) or vertical stiffness and (3) assessed validity or reliability. Meta-analyses were performed for a pooled validity estimate when (1) studies reported means and standard deviation for variables derived from the IMU and criterion (2) used the same IMU placement and (3) determined validity at a comparable running velocity (≤ 1 m·s^-1 difference).

RESULTS

Thirty-nine articles were included, where placement varied between the foot, tibia, hip, sacrum, lumbar spine (LS), torso and thoracic spine (TS). Initial contact, toe-off, contact time (CT), flight time (FT), step time, stride time, swing time, step frequency (SF), step length (SL), stride length, peak vertical and resultant GRF and vertical stiffness were analysed. Four variables (CT, FT, SF and SL) were meta-analysed, where CT was compared between the foot, tibia and LS placements and SF was compared between foot and LS. Foot placement data were meta-analysed for FT and SL. All data are the mean difference (MD [95%CI]). No significant difference was observed for any site compared to the criterion for CT (foot: - 11.47 ms [- 45.68, 22.74], p = 0.43; tibia: 22.34 ms [- 18.59, 63.27], p = 0.18; LS: - 48.74 ms [- 120.33, 22.85], p = 0.12), FT (foot: 11.93 ms [- 8.88, 32.74], p = 0.13), SF (foot: 0.45 step·min^-1 [- 1.75, 2.66], p = 0.47; LS: - 3.45 step·min^-1 [- 16.28, 9.39], p = 0.37) and SL (foot: 0.21 cm [- 1.76, 2.18], p = 0.69). Reliable derivations of CT (coefficient of variation [CV] < 9.9%), FT (CV < 11.6%) and SF (CV < 4.4%) were shown using foot- and LS-worn IMUs, while the CV was < 7.8% for foot-determined stride time, SL and stride length. Vertical GRF was reliable from the LS (CV = 4.2%) and TS (CV = 3.3%) using a spring-mass model, while vertical stiffness was moderately (r = 0.66) and nearly perfectly (r = 0.98) correlated with criterion measures from the TS.

CONCLUSION

Placement of IMUs on the foot, tibia and LS is suitable to derive valid and reliable stride data, suggesting measurement site may not be a critical factor. However, evidence regarding the ability to accurately detect stride events from the TS is unclear and this warrants further investigation.

Indirect Estimation of Vertical Ground Reaction Force from a Body-Mounted INS/GPS Using Machine Learning

Vertical ground reaction force (vGRF) can be measured by force plates or instrumented treadmills, but their application is limited to indoor environments. Insoles remove this restriction but suffer from low durability (several hundred hours). Therefore, interest in the indirect estimation of vGRF using inertial measurement units and machine learning techniques has increased. This paper presents a methodology for indirectly estimating vGRF and other features used in gait analysis from measurements of a wearable GPS-aided inertial navigation system (INS/GPS) device. A set of 27 features was extracted from the INS/GPS data. Feature analysis showed that six of these features suffice to provide precise estimates of 11 different gait parameters. Bagged ensembles of regression trees were then trained and used for predicting gait parameters for a dataset from the test subject from whom the training data were collected and for a dataset from a subject for whom no training data were available. The prediction accuracies for the latter were significantly worse than for the first subject but still sufficiently good. K-nearest neighbor (KNN) and long short-term memory (LSTM) neural networks were then used for predicting vGRF and ground contact times. The KNN yielded a lower normalized root mean square error than the neural network for vGRF predictions but cannot detect new patterns in force curves.

Predicting gait events from tibial acceleration in rearfoot running: A structured machine learning approach

BACKGROUND

Gait event detection of the initial contact and toe off is essential for running gait analysis, allowing the derivation of parameters such as stance time. Heuristic-based methods exist to estimate these key gait events from tibial accelerometry. However, these methods are tailored to very specific acceleration profiles, which may offer complications when dealing with larger data sets and inherent biological variability.

RESEARCH QUESTION

Can a structured machine learning approach achieve a more accurate prediction of running gait event timings from tibial accelerometry, compared to the previously utilised heuristic approaches?

METHODS

Force-based event detection acted as the criterion measure in order to assess the accuracy, repeatability and sensitivity of the predicted gait events. 3D tibial acceleration and ground reaction force data from 93 rearfoot runners were captured. A heuristic method and two structured machine learning methods were employed to derive initial contact, toe off and stance time from tibial acceleration signals.

RESULTS

Both a structured perceptron model (median absolute error of stance time estimation: 10.00 ± 8.73 ms) and a structured recurrent neural network model (median absolute error of stance time estimation: 6.50 ± 5.74 ms) significantly outperformed the existing heuristic approach (median absolute error of stance time estimation: 11.25 ± 9.52 ms). Thus, results indicate that a structured recurrent neural network machine learning model offers the most accurate and consistent estimation of the gait events and its derived stance time during level overground running.

SIGNIFICANCE

The machine learning methods seem less affected by intra- and inter-subject variation within the data, allowing for accurate and efficient automated data output during rearfoot overground running. Furthermore offering possibilities for real-time monitoring and biofeedback during prolonged measurements, even outside the laboratory.

Classification of runners' performance levels with concurrent prediction of biomechanical parameters using data from inertial measurement units

Identification of runner's performance level is critical to coaching, performance enhancement and injury prevention. Machine learning techniques have been developed to measure biomechanical parameters with body-worn inertial measurement unit (IMU) sensors. However, a robust method to classify runners is still unavailable. In this paper, we developed two models to classify running performance and predict biomechanical parameters of 30 subjects. We named the models RunNet-CNN and RunNet-MLP based on their architectures: convolutional neural network (CNN) and multilayer perceptron (MLP), respectively. In addition, we examined two validation approaches, subject-wise (leave-one-subject-out) and record-wise. RunNet-MLP classified runner's performance levels with an overall accuracy of 97.1%. Our results also showed that RunNet-CNN outperformed RunNet-MLP and gradient boosting decision tree in predicting biomechanical parameters. RunNet-CNN showed good agreement (R² > 0.9) with the ground-truth reference on biomechanical parameters. The prediction accuracy for the record-wise method was better than the subject-wise method regardless of biomechanical parameters or models. Our findings showed the viability of using IMUs to produce reliable prediction of runners' performance levels and biomechanical parameters.

Estimation of Three-Dimensional Lower Limb Kinetics Data during Walking Using Machine Learning from a Single IMU Attached to the Sacrum

Kinetics data such as ground reaction forces (GRFs) are commonly used as indicators for rehabilitation and sports performance; however, they are difficult to measure with convenient wearable devices. Therefore, researchers have attempted to estimate accurately unmeasured kinetics data with artificial neural networks (ANNs). Because the inputs to an ANN affect its performance, they must be carefully selected. The GRF and center of pressure (CoP) have a mechanical relationship with the center of mass (CoM) in the three dimensions (3D). This biomechanical characteristic can be used to establish an appropriate input and structure of an ANN. In this study, an ANN for estimating gait kinetics with a single inertial measurement unit (IMU) was designed; the kinematics of the IMU placed on the sacrum as a proxy for the CoM kinematics were applied based on the 3D spring mechanics. The walking data from 17 participants walking at various speeds were used to train and validate the ANN. The estimated 3D GRF, CoP trajectory, and joint torques of the lower limbs were reasonably accurate, with normalized root-mean-square errors (NRMSEs) of 6.7% to 15.6%, 8.2% to 20.0%, and 11.4% to 24.1%, respectively. This result implies that the biomechanical characteristics can be used to estimate the complete three-dimensional gait data with an ANN model and a single IMU.

Real-Time Vertical Ground Reaction Force Estimation in a Unified Simulation Framework Using Inertial Measurement Unit Sensors

Recent advances in computational technology have enabled the use of model-based simulation with real-time motion tracking to estimate ground reaction forces during gait. We show here that a biomechanical-based model including a foot-ground contact can reproduce measured ground reaction forces using inertial measurement unit data during single-leg support, single-support jump, side to side jump, jogging, and skipping. The framework is based on our previous work on integrating the OpenSim musculoskeletal models with the Unity environment. The validation was performed on a single subject performing several tasks that involve the lower extremity. The novelty of this paper includes the integration and real-time tracking of inertial measurement unit data in the current framework, as well as the estimation of contact forces using biologically based musculoskeletal models. The RMS errors of tracking the vertical ground reaction forces are 0.027 bodyweight, 0.174 bodyweight, 0.173 bodyweight, 0.095 bodyweight, and 0.10 bodyweight for single-leg support, single-support jump, side to side jump, jogging, and skipping, respectively. The average RMS error for all tasks and trials is 0.112 bodyweight. This paper provides a computational framework for further applications in whole-body human motion analysis.

Artificial Neural Networks in Motion Analysis-Applications of Unsupervised and Heuristic Feature Selection Techniques

The use of machine learning to estimate joint angles from inertial sensors is a promising approach to in-field motion analysis. In this context, the simplification of the measurements by using a small number of sensors is of great interest. Neural networks have the opportunity to estimate joint angles from a sparse dataset, which enables the reduction of sensors necessary for the determination of all three-dimensional lower limb joint angles. Additionally, the dimensions of the problem can be simplified using principal component analysis. Training a long short-term memory neural network on the prediction of 3D lower limb joint angles based on inertial data showed that three sensors placed on the pelvis and both shanks are sufficient. The application of principal component analysis to the data of five sensors did not reveal improved results. The use of longer motion sequences compared to time-normalised gait cycles seems to be advantageous for the prediction accuracy, which bridges the gap to real-time applications of long short-term memory neural networks in the future.

Estimating Vertical Ground Reaction Force during Walking Using a Single Inertial Sensor

The vertical ground reaction force (vGRF) and its passive and active peaks are important gait parameters and of great relevance for musculoskeletal injury analysis and prevention, the detection of gait abnormities, and the evaluation of lower-extremity prostheses. Most currently available methods to estimate the vGRF require a force plate. However, in real-world scenarios, gait monitoring would not be limited to a laboratory setting. This paper reports a novel solution using machine learning algorithms to estimate the vGRF and the timing and magnitude of its peaks from data collected by a single inertial measurement unit (IMU) on one of the lower limb locations. Nine volunteers participated in this study, walking on a force plate-instrumented treadmill at various speeds. Four IMUs were worn on the foot, shank, distal thigh, and proximal thigh, respectively. A random forest model was employed to estimate the vGRF from data collected by each of the IMUs. We evaluated the performance of the models against the gold standard measurement of the vGRF generated by the treadmill. The developed model achieved a high accuracy with a correlation coefficient, root mean square error, and normalized root mean square error of 1.00, 0.02 body weight (BW), and 1.7% in intra-participant testing, and 0.97, 0.10 BW, and 7.15% in inter-participant testing, respectively, for the shank location. The difference between the reference and estimated passive force peak values was 0.02 BW and 0.14 BW with a delay of −0.14% and 0.57% of stance duration for the intra- and inter-participant testing, respectively; the difference between the reference and estimated active force peak values was 0.02 BW and 0.08 BW with a delay of 0.45% and 1.66% of stance duration for the intra- and inter-participant evaluation, respectively. We concluded that vertical ground reaction force can be estimated using only a single IMU via machine learning algorithms. This research sheds light on the development of a portable wearable gait monitoring system reporting the real-time vGRF in real-life scenarios.

Accelerometer-based prediction of skeletal mechanical loading during walking in normal weight to severely obese subjects

There is no objective way to monitor mechanical loading characteristics during exercise for bone health improvement. We developed accelerometry-based equations to predict ground reaction force (GRF) and loading rate (LR) in normal weight to severely obese subjects. Equations developed had a high and moderate accuracy for GRF and LR prediction, respectively, thereby representing an accessible way to determine mechanical loading characteristics in clinical settings.

INTRODUCTION

There is no way to objectively prescribe and monitor exercise for bone health improvement in obese patients based on mechanical loading characteristics. We aimed to develop accelerometry-based equations to predict peak ground reaction forces (pGRFs) and peak loading rate (pLR) on normal weight to severely obese subjects.

METHODS

Sixty-four subjects (45 females; 84.6 ± 21.7 kg) walked at different speeds (2-6 km·h^-1) on a force plate-equipped treadmill while wearing accelerometers at lower back and hip. Regression equations were developed to predict pGRF and pLR from accelerometry data. Leave-one-out cross-validation was used to calculate prediction accuracy and Bland-Altman plots. Actual and predicted values at different speeds were compared by repeated measures ANOVA.

RESULTS

Body mass and peak acceleration were included for pGRF prediction and body mass and peak acceleration transient rate for pLR prediction. All pGRF equation coefficients of determination were above 0.89, a good agreement between actual and predicted pGRFs, with a mean absolute percent error (MAPE) below 6.7%. No significant differences were observed between actual and predicted pGRFs at each walking speed. Accuracy indices from our equations were better than previously developed equations for normal weight subjects, namely a MAPE approximately 3 times smaller. All pLR prediction equations presented a lower accuracy compared to those developed to predict pGRF.

CONCLUSION

Walking pGRF and pLR in normal weight to severely obese subjects can be predicted with moderate to high accuracy by accelerometry-based equations, representing an easy and accessible way to determine mechanical loading characteristics in clinical settings.

Tibial Acceleration-Based Prediction of Maximal Vertical Loading Rate During Overground Running: A Machine Learning Approach

Ground reaction forces are often used by sport scientists and clinicians to analyze the mechanical risk-factors of running related injuries or athletic performance during a running analysis. An interesting ground reaction force-derived variable to track is the maximal vertical instantaneous loading rate (VILR). This impact characteristic is traditionally derived from a fixed force platform, but wearable inertial sensors nowadays might approximate its magnitude while running outside the lab. The time-discrete axial peak tibial acceleration (APTA) has been proposed as a good surrogate that can be measured using wearable accelerometers in the field. This paper explores the hypothesis that applying machine learning to time continuous data (generated from bilateral tri-axial shin mounted accelerometers) would result in a more accurate estimation of the VILR. Therefore, the purpose of this study was to evaluate the performance of accelerometer-based predictions of the VILR with various machine learning models trained on data of 93 rearfoot runners. A subject-dependent gradient boosted regression trees (XGB) model provided the most accurate estimates (mean absolute error: 5.39 ± 2.04 BW⋅s^–1, mean absolute percentage error: 6.08%). A similar subject-independent model had a mean absolute error of 12.41 ± 7.90 BW⋅s^–1 (mean absolute percentage error: 11.09%). All of our models had a stronger correlation with the VILR than the APTA (p < 0.01), indicating that multiple 3D acceleration features in a learning setting showed the highest accuracy in predicting the lab-based impact loading compared to APTA.

Effects of segment masses and cut-off frequencies on the estimation of vertical ground reaction forces in running

The purpose of this study is to examine the effect of the body's mass distribution to segments and the filtering of kinematic data on the estimation of vertical ground reaction forces from positional data. A public dataset of raw running biomechanics was used for the purposes of the analysis, containing recordings of twenty-eight competitive or elite athletes running on an instrumented treadmill at three different speeds. A grid-search on half of the trials was employed to seek the values of the parameters that optimise the approximation of biomechanical loads. Two-way ANOVAs were then conducted to examine the significance of the parameterised factors in the modelled waveforms. The reserved recordings were used to validate the predictive accuracy of the model. The cut-off filtering frequencies of the pelvis and thigh markers were correlated to running speed and heel-strike patterns, respectively. Optimal segment masses were in agreement with standardised literature reported values. Root mean square errors for slow running (2.5 m/s) were on average equal to 0.1 (body weight normalized). Errors increased with running speeds to 0.13 and 0.18 for 3.5 m/s and 4.5 m/s, respectively. This study accurately estimated vertical ground reaction forces for slow-paced running by only considering the kinematics of the pelvis and thighs. Future studies should consider configuring the filtering of kinematic inputs based on the location of markers and type of running.

Prediction of Lower Limb Kinetics and Kinematics during Walking by a Single IMU on the Lower Back Using Machine Learning

Recent studies have reported the application of artificial neural network (ANN) techniques on data of inertial measurement units (IMUs) to predict ground reaction forces (GRFs), which could serve as quantitative indicators of sports performance or rehabilitation. The number of IMUs and their measurement locations are often determined heuristically, and the rationale underlying the selection of these parameter values is not discussed. Using the dynamic relationship between the center of mass (CoM), the GRFs and joint kinetics, we propose the CoM as a single measurement location with which to predict the dynamic data of the lower limbs, using an ANN. Data from seven subjects walking on a treadmill at various speeds were collected from a single IMU worn near the sacrum. The data was segmented by step and numerically processed for integration. Six segment angles of the stance and swing leg, three joint torques, and two GRFs were estimated from the kinematics of the CoM measured from a single IMU sensor, with fair accuracy. These results indicate the importance of the CoM as a dynamic determinant of multi-segment kinetics during walking. The tradeoff between data quantity and wearable convenience can be solved by utilizing a machine learning algorithm based on the dynamic characteristics of human walking.

Machine learning approach to predict center of pressure trajectories in a complete gait cycle: a feedforward neural network vs. LSTM network

Center of pressure (COP) trajectories of human can maintain regulation of forward progression and stability of lateral sway during walking. The insole pressure system can only detect COP trajectories of each foot during single stance. In this study, we developed artificial neural network models that could present COP trajectories in an integrated coordinate system during a complete gait cycle using pressure information of the insole system. A feed forward artificial neural network (FFANN) and a long short-term memory (LSTM) model were developed. For FFANN, among 198 pressure sensors from Pedar-X insoles, proper input variables were selected using sequential forward selection to reduce input dimension. The LSTM model used all 198 signals as inputs because of its self-learning characteristic. As results of cross-validation, the FFANN model showed correlation coefficients of 0.98-0.99 and 0.93-0.95 in anterior/posterior and medial/lateral directions, respectively. For the LSTM model, correlation coefficients were similar to those of FFANN. However, the relative root mean square error (12.5%) of the FFANN model was higher than that (9.8%) of the LSTM model in medial/lateral direction (p = 0.03). This study can be used for quantitative evaluation of clinical diagnosis and rehabilitation status for patient with various diseases through further training using varied databases. Graphical abstract Architectures of neural networks developed in this study (a feed forward artificial neural network; b LSTM network).

Estimation of Vertical Ground Reaction Force Using Low-Cost Insole With Force Plate-Free Learning From Single Leg Stance and Walking

For the evaluation of pathological gait, a machine learning-based estimation of the vertical ground reaction force (vGRF) using a low-cost insole is proposed as an alternative to costly force plates. However, learning a model for estimation still relies on the use of force plates, which is not accessible in small clinics and individuals. Therefore, this paper presents a force plate-free learning from a single leg stance (SLS) and natural walking measured only by the insoles. This method used a linear least squares regression that fits insole measurements during SLS to body weight in order to learn a model to estimate vGRF during walking. Constraints were added to the regression so that vGRF estimates during walking were of proper magnitude, and the constraint bounds were newly defined as a linear function of stance duration. Moreover, a lower bound for the estimated vGRF in mid-stance was added to the constraints to enhance estimation accuracy. The vGRF estimated by the proposed method was compared with force platforms for 4 healthy young adults and 13 elderly adults including patients with mild osteoarthritis, knee pain, and valgus hallux. Through the experiments, the proposed learning method had a normalized root mean squared error under 10% for healthy young and elderly adults with stance durations within a certain range (600-800 ms). From these results, the validity of the proposed learning method was verified for various users requiring assessment in the field of medicine and healthcare.

Single Inertial Sensor-Based Neural Networks to Estimate COM-COP Inclination Angle During Walking

A biomechanical understanding of gait stability is needed to reduce falling risk. As a typical parameter, the COM-COP (center of mass-center of pressure) inclination angle (IA) could provide valuable insight into postural control and balance recovery ability. In this study, an artificial neural network (ANN) model was developed to estimate COM-COP IA based on signals using an inertial sensor. Also, we evaluated how different types of ANN and the cutoff frequency of the low-pass filter applied to input signals could affect the accuracy of the model. An inertial measurement unit (IMU) including an accelerometer, gyroscope, and magnetometer sensors was fabricated as a prototype. The COM-COP IA was calculated using a 3D motion analysis system including force plates. In order to predict the COM-COP IA, a feed-forward ANN and long-short term memory (LSTM) network was developed. As a result, the feed-forward ANN showed a relative root-mean-square error (rRMSE) of 15% while the LSTM showed an improved accuracy of 9% rRMSE. Additionally, the LSTM displayed a stable accuracy regardless of the cutoff frequency of the filter applied to the input signals. This study showed that estimating the COM-COP IA was possible with a cheap inertial sensor system. Furthermore, the neural network models in this study can be implemented in systems to monitor the balancing ability of the elderly or patients with impaired balancing ability.

Insole-Based Estimation of Vertical Ground Reaction Force Using One-Step Learning With Probabilistic Regression and Data Augmentation

An insole-based estimation of the vertical ground reaction force (vGRF) is proposed as an alternative to costly force plates for the evaluation of pathological gait. However, machine learning techniques for estimation still rely on the use of force plates. Moreover, measuring plural walking steps in order to prevent overfitting induces fall risks and physically taxes the patients. Therefore, this paper presents an accessible and efficient learning scheme for the insole-based estimation of vGRF. In this system, we employ a low-cost scale as an alternative to force plates. Then, we use Gaussian process regression (GPR) to learn a model in order to estimate vGRF without overfitting of small-sized data sets corrupted by measurement errors and noise of the devices. In addition, we propose a "one-step learning" scheme based on a probabilistic data augmentation. This approach augments actual measurements of a minimum (just one) walking step to a virtual data set for plural steps by considering their typical variability between steps. In experiments, the GPR models learned from two walking steps estimated vGRF with mean errors of 8% or under for entire/local magnitudes. Moreover, the learning from one step with probabilistic augmentation enhanced the estimation accuracy.

Continuous Analysis of Running Mechanics by Means of an Integrated INS/GPS Device

This paper describes a single body-mounted sensor that integrates accelerometers, gyroscopes, compasses, barometers, a GPS receiver, and a methodology to process the data for biomechanical studies. The sensor and its data processing system can accurately compute the speed, acceleration, angular velocity, and angular orientation at an output rate of 400 Hz and has the ability to collect large volumes of ecologically-valid data. The system also segments steps and computes metrics for each step. We analyzed the sensitivity of these metrics to changing the start time of the gait cycle. Along with traditional metrics, such as cadence, speed, step length, and vertical oscillation, this system estimates ground contact time and ground reaction forces using machine learning techniques. This equipment is less expensive and cumbersome than the currently used alternatives: Optical tracking systems, in-shoe pressure measurement systems, and force plates. Another advantage, compared to existing methods, is that natural movement is not impeded at the expense of measurement accuracy. The proposed technology could be applied to different sports and activities, including walking, running, motion disorder diagnosis, and geriatric studies. In this paper, we present the results of tests in which the system performed real-time estimation of some parameters of walking and running which are relevant to biomechanical research. Contact time and ground reaction forces computed by the neural network were found to be as accurate as those obtained by an in-shoe pressure measurement system.