Estimating Biomechanical Time-Series with Wearable Sensors: A Systematic Review of Machine Learning Techniques

The AI Future of Emergency Medicine

In the coming years, artificial intelligence (AI) and machine learning will likely give rise to profound changes in the field of emergency medicine, and medicine more broadly. This article discusses these anticipated changes in terms of 3 overlapping yet distinct stages of AI development. It reviews some fundamental concepts in AI and explores their relation to clinical practice, with a focus on emergency medicine. In addition, it describes some of the applications of AI in disease diagnosis, prognosis, and treatment, as well as some of the practical issues that they raise, the barriers to their implementation, and some of the legal and regulatory challenges they create.

Estimating 3D kinematics and kinetics from virtual inertial sensor data through musculoskeletal movement simulations

Portable measurement systems using inertial sensors enable motion capture outside the lab, facilitating longitudinal and large-scale studies in natural environments. However, estimating 3D kinematics and kinetics from inertial data for a comprehensive biomechanical movement analysis is still challenging. Machine learning models or stepwise approaches performing Kalman filtering, inverse kinematics, and inverse dynamics can lead to inconsistencies between kinematics and kinetics. We investigated the reconstruction of 3D kinematics and kinetics of arbitrary running motions from inertial sensor data using optimal control simulations of full-body musculoskeletal models. To evaluate the feasibility of the proposed method, we used marker tracking simulations created from optical motion capture data as a reference and for computing virtual inertial data such that the desired solution was known exactly. We generated the inertial tracking simulations by formulating optimal control problems that tracked virtual acceleration and angular velocity while minimizing effort without requiring a task constraint or an initial state. To evaluate the proposed approach, we reconstructed three trials each of straight running, curved running, and a v-cut of 10 participants. We compared the estimated inertial signals and biomechanical variables of the marker and inertial tracking simulations. The inertial data was tracked closely, resulting in low mean root mean squared deviations for pelvis translation (≤20.2 mm), angles (≤1.8 deg), ground reaction forces (≤1.1 BW%), joint moments (≤0.1 BWBH%), and muscle forces (≤5.4 BW%) and high mean coefficients of multiple correlation for all biomechanical variables ( ≥ 0.99 ) . Accordingly, our results showed that optimal control simulations tracking 3D inertial data could reconstruct the kinematics and kinetics of individual trials of all running motions. The simulations led to mutually and dynamically consistent kinematics and kinetics, which allows researching causal chains, for example, to analyze anterior cruciate ligament injury prevention. Our work proved the feasibility of the approach using virtual inertial data. When using the approach in the future with measured data, the sensor location and alignment on the segment must be estimated, and soft-tissue artifacts are potential error sources. Nevertheless, we demonstrated that optimal control simulation tracking inertial data is highly promising for estimating 3D kinematics and kinetics for a comprehensive biomechanical analysis.

3D gait analysis in children using wearable sensors: feasibility of predicting joint kinematics and kinetics with personalized machine learning models and inertial measurement units

Introduction: Children's walking patterns evolve with age, exhibiting less repetitiveness at a young age and more variability than adults. Three-dimensional gait analysis (3DGA) is crucial for understanding and treating lower limb movement disorders in children, traditionally performed using Optical Motion Capture (OMC). Inertial Measurement Units (IMUs) offer a cost-effective alternative to OMC, although challenges like drift errors persist. Machine learning (ML) models can mitigate these issues in adults, prompting an investigation into their applicability to a heterogeneous pediatric population. This study aimed at 1) quantifying personalized and generalized ML models' performance for predicting gait time series in typically developed (TD) children using IMUs data, 2) Comparing random forest (RF) and convolutional neural networks (CNN) models' performance, 3) Finding the optimal number of IMUs required for accurate predictions. Methodology: Seventeen TD children, aged 6 to 15, participated in data collection involving OMC, force plates, and IMU sensors. Joint kinematics and kinetics (targets) were computed from OMC and force plates' data using OpenSim. Tsfresh, a Python package, extracted features from raw IMU data. Each target's ten most important features were input in the development of personalized and generalized RF and CNN models. This procedure was initially conducted with 7 IMUs placed on all lower limb segments and then performed using only two IMUs on the feet. Results: Findings suggested that the RF and CNN models demonstrated comparable performance. RF predicted joint kinematics with a 9.5% and 19.9% NRMSE for personalized and generalized models, respectively, and joint kinetics with an NRMSE of 10.7% for personalized and 15.2% for generalized models in TD children. Personalized models provided accurate estimations from IMU data in children, while generalized models lacked accuracy due to the limited dataset. Furthermore, reducing the number of IMUs from 7 to 2 did not affect the results, and the performance remained consistent. Discussion: This study proposed a promising personalized approach for gait time series prediction in children, involving an RF model and two IMUs on the feet.

A large clinical trial to improve well-being during the transition to college using wearables: The lived experiences measured using rings study

INTRODUCTION

The transition to college is a period of elevated risk for a range of mental health conditions. Although colleges and universities strive to provide mental health support to their students, the high demand for these services makes it difficult to provide scalable, cost-effective solutions.

OBJECTIVE

To address these issues, the present study aims to compare the efficacy of three different treatments using a large cohort of 600 students transitioning to college. Interventions were selected based on their potential for generalizability and cost-effectiveness on college campuses.

METHODS

The study is a Phase II parallel-group, four-arm, randomized controlled trial with 1:1 allocation that will assign 600 participants to one (n = 150 per condition) of four arms: 1) group-based therapy, 2) physical activity program, 3) nature experiences, or 4) weekly assessment condition as a control group. Physiological data will be collected from all participants using a wearable device to develop algorithmic mental and physical health functioning predictions. Once recruitment is complete, modeling strategies will be used to evaluate the outcomes and effectiveness of each intervention.

DISCUSSION

The findings of this study will provide evidence as to the benefits of implementing scalable and proactive interventions using technology with the goal of improving the well-being and success of new college students.

Challenges and advances in the use of wearable sensors for lower extremity biomechanics

The use of wearable sensors for the collection of lower extremity biomechanical data is increasing in popularity, in part due to the ease of collecting data and the ability to capture movement outside of traditional biomechanics laboratories. Consequently, an increasing number of researchers are facing the challenges that come with utilizing the data captured by wearable sensors. These challenges include identifying/calculating meaningful measures from unfamiliar data types (measures of acceleration and angular velocity instead of positions and joint angles), defining sensor-to-segment alignments for calculating traditional biomechanics metrics, using reduced sensor sets and machine learning to predict unmeasured signals, making decisions about when and how to make algorithms freely available, and developing or replicating methods to perform basic processing tasks such as recognizing activities of interest or identifying gait events. In this perspective article, we present our own approaches to common challenges in lower extremity biomechanics research using wearable sensors and share our perspectives on approaching several of these challenges. We present these perspectives with examples that come mostly from gait research, but many of the concepts also apply to other contexts where researchers may use wearable sensors. Our goal is to introduce common challenges to new users of wearable sensors, and to promote dialogue amongst experienced users towards best practices.

Estimating 3D ground reaction forces in running using three inertial measurement units

To understand the mechanisms causing running injuries, it is crucial to get insights into biomechanical loading in the runners' environment. Ground reaction forces (GRFs) describe the external forces on the body during running, however, measuring these forces is usually only possible in a gait laboratory. Previous studies show that it is possible to use inertial measurement units (IMUs) to estimate vertical forces, however, forces in anterior-posterior direction play an important role in the push-off. Furthermore, to perform an inverse dynamics approach, for modelling tissue specific loads, 3D GRFs are needed as input. Therefore, the goal of this work was to estimate 3D GRFs using three inertial measurement units. Twelve rear foot strike runners did nine trials at three different velocities (10, 12 and 14 km/h) and three stride frequencies (preferred and preferred ± 10%) on an instrumented treadmill. Then, data from IMUs placed on the pelvis and lower legs were used as input for artificial neural networks (ANNs) to estimate 3D GRFs. Additionally, estimated vertical GRF from a physical model was used as input to create a hybrid machine learning model. Using different splits in validation and training data, different ANNs were fitted and assembled into an ensemble model. Leave-one-subject-out cross-validation was used to validate the models. Performance of the machine learning, hybrid machine learning and a physical model were compared. The estimated vs. measured GRF for the hybrid model had a RMSE normalized over the full range of values of 10.8, 7.8 and 6.8% and a Pearson correlation coefficient of 0.58, 0.91, 0.97 for the mediolateral direction, posterior-anterior and vertical direction respectively. Performance for the three compared models was similar. The ensemble models showed higher model accuracy compared to the ensemble-members. This study is the first to estimate 3D GRF during continuous running from IMUs and shows that it is possible to estimate GRF in posterior-anterior and vertical direction, making it possible to estimate these forces in the outdoor setting. This step towards quantification of biomechanical load in the runners' environment is helpful to gain a better understanding of the development of running injuries.

A Wearable Real-time Kinematic and Kinetic Measurement Sensor Setup for Human Locomotion

Current laboratory-based setups (optical marker cameras + force plates) for human motion measurement require participants to stay in a constrained capture region which forbids rich movement types. This study established a fully wearable system, based on commercially available sensors (inertial measurement units + pressure insoles) that can measure both kinematic and kinetic motion data simultaneously and support wireless frame-by-frame streaming. In addition, its capability and accuracy were tested against a conventional laboratory-based setup. An experiment was conducted, with 9 participants wearing the wearable measurement system and performing 13 daily motion activities, from slow walking to fast running, together with vertical jump, squat, lunge and single-leg landing, inside the capture space of the laboratory-based motion capture system. The recorded sensor data were post-processed to obtain joint angles, ground reaction forces (GRFs), and joint torques (via multi-body inverse dynamics). Compared to the laboratory-based system, the established wearable measurement system can measure accurate information of all lower limb joint angles (Pearson's r = 0.929), vertical GRFs (Pearson's r = 0.954), and ankle joint torques (Pearson's r = 0.917). Center of pressure (CoP) in the anterior-posterior direction and knee joint torques were fairly matched (Pearson's r = 0.683 and 0.612, respectively). Calculated hip joint torques and measured medial-lateral CoP did not match with the laboratory-based system (Pearson's r = 0.21 and 0.47, respectively). Furthermore, both raw and processed datasets are openly accessible (https://doi.org/10.5281/zenodo.6457662). Documentation, data processing codes, and guidelines to establish the real-time wearable kinetic measurement system are also shared (https://github.com/HuaweiWang/WearableMeasurementSystem).

Estimation of ground reaction force waveforms during fixed pace running outside the laboratory

In laboratory experiments, biomechanical data collections with wearable technologies and machine learning have been promising. Despite the development of lightweight portable sensors and algorithms for the identification of gait events and estimation of kinetic waveforms, machine learning models have yet to be used to full potential. We propose the use of a Long Short Term Memory network to map inertial data to ground reaction force data gathered in a semi-uncontrolled environment. Fifteen healthy runners were recruited for this study, with varied running experience: novice to highly trained runners (<15 min 5 km race), and ages ranging from 18 to 64 years old. Force sensing insoles were used to measure normal foot-shoe forces, providing the standard for identification of gait events and measurement of kinetic waveforms. Three inertial measurement units (IMUs) were mounted to each participant, two bilaterally on the dorsal aspect of the foot and one clipped to the back of each participant's waistband, approximating their sacrum. Data input into the Long Short Term Memory network were from the three IMUs and output were estimated kinetic waveforms, compared against the standard of the force sensing insoles. The range of RMSE for each stance phase was from 0.189-0.288 BW, which is similar to multiple previous studies. Estimation of foot contact had an r ² = 0.795. Estimation of kinetic variables varied, with peak force presenting the best output with an r ² = 0.614. In conclusion, we have shown that at controlled paces over level ground a Long Short Term Memory network can estimate 4 s temporal windows of ground reaction force data across a range of running speeds.

Estimation of gait events and kinetic waveforms with wearable sensors and machine learning when running in an unconstrained environment

Wearable sensors and machine learning algorithms are becoming a viable alternative for biomechanical analysis outside of the laboratory. The purpose of this work was to estimate gait events from inertial measurement units (IMUs) and utilize machine learning for the estimation of ground reaction force (GRF) waveforms. Sixteen healthy runners were recruited for this study, with varied running experience. Force sensing insoles were used to measure normal foot-shoe forces, providing a proxy for vertical GRF and a standard for the identification of gait events. Three IMUs were mounted on each participant, two bilaterally on the dorsal aspect of each foot and one clipped to the back of each participant's waistband, approximating their sacrum. Participants also wore a GPS watch to record elevation and velocity. A Bidirectional Long Short Term Memory Network (BD-LSTM) was used to estimate GRF waveforms from inertial waveforms. Gait event estimation from both IMU data and machine learning algorithms led to accurate estimations of contact time. The GRF magnitudes were generally underestimated by the machine learning algorithm when presented with data from a novel participant, especially at faster running speeds. This work demonstrated that estimation of GRF waveforms is feasible across a range of running velocities and at different grades in an uncontrolled environment.

Perspective on "in the wild" movement analysis using machine learning

Recent advances in wearable sensing and machine learning have created ample opportunities for "in the wild" movement analysis in sports, since the combination of both enables real-time feedback to be provided to athletes and coaches, as well as long-term monitoring of movements. The potential for real-time feedback is useful for performance enhancement or technique analysis, and can be achieved by training efficient models and implementing them on dedicated hardware. Long-term monitoring of movement can be used for injury prevention, among others. Such applications are often enabled by training a machine learned model from large datasets that have been collected using wearable sensors. Therefore, in this perspective paper, we provide an overview of approaches for studies that aim to analyze sports movement "in the wild" using wearable sensors and machine learning. First, we discuss how a measurement protocol can be set up by answering six questions. Then, we discuss the benefits and pitfalls and provide recommendations for effective training of machine learning models from movement data, focusing on data pre-processing, feature calculation, and model selection and tuning. Finally, we highlight two application domains where "in the wild" data recording was combined with machine learning for injury prevention and technique analysis, respectively.

Prediction of continuous and discrete kinetic parameters in horses from inertial measurement units data using recurrent artificial neural networks

Vertical ground reaction force (GRFz) measurements are the best tool for assessing horses' weight-bearing lameness. However, collection of these data is often impractical for clinical use. This study evaluates GRFz predicted using data from body-mounted IMUs and long short-term memory recurrent neural networks (LSTM-RNN). Twenty-four clinically sound horses, equipped with IMUs on the upper-body (UB) and each limb, walked and trotted on a GRFz measuring treadmill (TiF). Both systems were time-synchronised. Data from randomly selected 16, 4, and 4 horses formed training, validation, and test datasets, respectively. LSTM-RNN with different input sets (All, Limbs, UB, Sacrum, or Withers) were trained to predict GRFz curves or peak-GRFz. Our models could predict GRFz shapes at both gaits with RMSE below 0.40 N.kg^-1. The best peak-GRFz values were obtained when extracted from the predicted curves by the all dataset. For both GRFz curves and peak-GRFz values, predictions made with the All or UB datasets were systematically better than with the Limbs dataset, showing the importance of including upper-body kinematic information for kinetic parameters predictions. More data should be gathered to confirm the usability of LSTM-RNN for GRFz predictions, as they highly depend on factors like speed, gait, and the presence of weight-bearing lameness.

Developing a method for quantifying hip joint angles and moments during walking using neural networks and wearables

Quantifying hip angles/moments during gait is critical for improving hip pathology diagnostic and treatment methods. Recent work has validated approaches combining wearables with artificial neural networks (ANNs) for cheaper, portable hip joint angle/moment computation. This study developed a Wearable-ANN approach for calculating hip joint angles/moments during walking in the sagittal/frontal planes with data from 17 healthy subjects, leveraging one shin-mounted inertial measurement unit (IMU) and a force-measuring insole for data capture. Compared to the benchmark approach, a two hidden layer ANN (n = 5 nodes per layer) achieved an average rRMSE = 15% and R²=0.85 across outputs, subjects and training rounds.

Automated classification of hand gestures using a wristband and machine learning for possible application in pill intake monitoring

BACKGROUND

Thanks to the increased interest towards health and lifestyle, a larger adoption in wearable devices for activity tracking is present among the general population. Wearable devices such as smart wristbands integrate inertial units, including accelerometers and gyroscopes, which can be utilised to perform automatic classification of hand gestures. This technology could also find an important application in automatic medication adherence monitoring. Accordingly, this study aims at comparing the performance of several Machine-Learning (ML) and Deep-Learning (DL) approaches for the automatic identification of hand gestures, with a specific focus on the drinking gesture, commonly associated to the action of oral intake of a pill-packed medication.

METHODS

A method to automatically recognize hand gestures in daily living is proposed in this work. The method relies on a commercially available wristband sensor (MetaMotionR, MbientLab Inc.) integrating tri-axial accelerometer and gyroscope. Both ML and DL algorithms were evaluated for both multi-gesture (drinking, eating, pouring water, opening a bottle, typing, answering a phone, combing hair, and cutting) and binary gesture (drinking versus other gestures) classification from wristband sensor signals. Twenty-two participants were involved in the experimental analysis, performing a 10 min acquisition in a laboratory setting. Leave-one-subject-out cross validation was performed for robust performance assessment.

RESULTS

The highest performance was achieved using a convolutional neural network with long- short term memory (CNN-LSTM), with a median f1-score of 90.5 [first quartile: 84.5; third quartile: 92.5]% and 92.5 [81.5;98.0]% for multi-gesture and binary classification, respectively.

CONCLUSIONS

Experimental results showed that hand gesture classification with ML/DL from wrist accelerometers and gyroscopes signals can be performed with reasonable accuracy in laboratory settings, paving the way for a new generation of medical devices for monitoring medical adherence.

Accelerometer-Based Identification of Fatigue in the Lower Limbs during Cyclical Physical Exercise: A Systematic Review

Physical exercise (PE) is beneficial for both physical and psychological health aspects. However, excessive training can lead to physical fatigue and an increased risk of lower limb injuries. In order to tailor training loads and durations to the needs and capacities of an individual, physical fatigue must be estimated. Different measurement devices and techniques (i.e., ergospirometers, electromyography, and motion capture systems) can be used to identify physical fatigue. The field of biomechanics has succeeded in capturing changes in human movement with optical systems, as well as with accelerometers or inertial measurement units (IMUs), the latter being more user-friendly and adaptable to real-world scenarios due to its wearable nature. There is, however, still a lack of consensus regarding the possibility of using biomechanical parameters measured with accelerometers to identify physical fatigue states in PE. Nowadays, the field of biomechanics is beginning to open towards the possibility of identifying fatigue state using machine learning algorithms. Here, we selected and summarized accelerometer-based articles that either (a) performed analyses of biomechanical parameters that change due to fatigue in the lower limbs or (b) performed fatigue identification based on features including biomechanical parameters. We performed a systematic literature search and analysed 39 articles on running, jumping, walking, stair climbing, and other gym exercises. Peak tibial and sacral acceleration were the most common measured variables and were found to significantly increase with fatigue (respectively, in 6/13 running articles and 2/4 jumping articles). Fatigue classification was performed with an accuracy between 78% and 96% and Pearson’s correlation with an RPE (rate of perceived exertion) between r = 0.79 and r = 0.95. We recommend future effort toward the standardization of fatigue protocols and methods across articles in order to generalize fatigue identification results and increase the use of accelerometers to quantify physical fatigue in PE.

Inertial Motion Capture-Based Wearable Systems for Estimation of Joint Kinetics: A Systematic Review

In biomechanics, joint kinetics has an important role in evaluating the mechanical load of the joint and understanding its motor function. Although an optical motion capture (OMC) system has mainly been used to evaluate joint kinetics in combination with force plates, inertial motion capture (IMC) systems have recently been emerging in joint kinetic analysis due to their wearability and ubiquitous measurement capability. In this regard, numerous studies have been conducted to estimate joint kinetics using IMC-based wearable systems. However, these have not been comprehensively addressed yet. Thus, the aim of this review is to explore the methodology of the current studies on estimating joint kinetic variables by means of an IMC system. From a systematic search of the literature, 48 studies were selected. This paper summarizes the content of the selected literature in terms of the (i) study characteristics, (ii) methodologies, and (iii) study results. The estimation methods of the selected studies are categorized into two types: the inverse dynamics-based method and the machine learning-based method. While these two methods presented different characteristics in estimating the kinetic variables, it was demonstrated in the literature that both methods could be applied with good performance for the kinetic analysis of joints in different daily activities.

Fusion of Wearable Kinetic and Kinematic Sensors to Estimate Triceps Surae Work during Outdoor Locomotion on Slopes

Muscle–tendon power output is commonly assessed in the laboratory through the work loop, a paired analysis of muscle force and length during a cyclic task. Work-loop analysis of muscle–tendon function in out-of-lab conditions has been elusive due to methodological limitations. In this work, we combined kinetic and kinematic measures from shear wave tensiometry and inertial measurement units, respectively, to establish a wearable system for estimating work and power output from the soleus and gastrocnemius muscles during outdoor locomotion. Across 11 healthy young adults, we amassed 4777 strides of walking on slopes from −10° to +10°. Results showed that soleus work scales with incline, while gastrocnemius work is relatively insensitive to incline. These findings agree with previous results from laboratory-based studies while expanding technological capabilities by enabling wearable analysis of muscle–tendon kinetics. Applying this system in additional settings and activities could improve biomechanical knowledge and evaluation of protocols in scenarios such as rehabilitation, device design, athletics, and military training.

Wearables-Only Analysis of Muscle and Joint Mechanics: An EMG-Driven Approach

Complex sensor arrays prohibit practical deployment of existing wearables-based algorithms for free-living analysis of muscle and joint mechanics. Machine learning techniques have been proposed as a potential solution, however, they are less interpretable and generalizable when compared to physics-based techniques. Herein, we propose a hybrid method utilizing inertial sensor- and electromyography (EMG)-driven simulation of muscle contraction to characterize knee joint and muscle mechanics during walking gait. Machine learning is used only to map a subset of measured muscle excitations to a full set thereby reducing the number of required sensors. We demonstrate the utility of the approach for estimating net knee flexion moment (KFM) as well as individual muscle moment and work during the stance phase of gait across nine unimpaired subjects. Across all subjects, KFM was estimated with 0.91%BW•H RMSE and strong correlations (r = 0.87) compared to ground truth inverse dynamics analysis. Estimates of individual muscle moments were strongly correlated (r = 0.81-0.99) with a reference EMG-driven technique using optical motion capture and a full set of electrodes as were estimates of muscle work (r = 0.88-0.99). Implementation of the proposed technique in the current work included instrumenting only three muscles with surface electrodes (lateral and medial gastrocnemius and vastus medialis) and both the thigh and shank segments with inertial sensors. These sensor locations permit instrumentation of a knee brace/sleeve facilitating a practically deployable mechanism for monitoring muscle and joint mechanics with performance comparable to the current state-of-the-art.

InverseMuscleNET: Alternative Machine Learning Solution to Static Optimization and Inverse Muscle Modeling

InverseMuscleNET, a machine learning model, is proposed as an alternative to static optimization for resolving the redundancy issue in inverse muscle models. A recurrent neural network (RNN) was optimally configured, trained, and tested to estimate the pattern of muscle activation signals. Five biomechanical variables (joint angle, joint velocity, joint acceleration, joint torque, and activation torque) were used as inputs to the RNN. A set of surface electromyography (EMG) signals, experimentally measured around the shoulder joint for flexion/extension, were used to train and validate the RNN model. The obtained machine learning model yields a normalized regression in the range of 88-91% between experimental data and estimated muscle activation. A sequential backward selection algorithm was used as a sensitivity analysis to discover the less dominant inputs. The order of most essential signals to least dominant ones was as follows: joint angle, activation torque, joint torque, joint velocity, and joint acceleration. The RNN model required 0.06 s of the previous biomechanical input signals and 0.01 s of the predicted feedback EMG signals, demonstrating the dynamic temporal relationships of the muscle activation profiles. The proposed approach permits a fast and direct estimation ability instead of iterative solutions for the inverse muscle model. It raises the possibility of integrating such a model in a real-time device for functional rehabilitation and sports evaluation devices with real-time estimation and tracking. This method provides clinicians with a means of estimating EMG activity without an invasive electrode setup.

Predicting continuous ground reaction forces from accelerometers during uphill and downhill running: a recurrent neural network solution

BACKGROUND

Ground reaction forces (GRFs) are important for understanding human movement, but their measurement is generally limited to a laboratory environment. Previous studies have used neural networks to predict GRF waveforms during running from wearable device data, but these predictions are limited to the stance phase of level-ground running. A method of predicting the normal (perpendicular to running surface) GRF waveform using wearable devices across a range of running speeds and slopes could allow researchers and clinicians to predict kinetic and kinematic variables outside the laboratory environment.

PURPOSE

We sought to develop a recurrent neural network capable of predicting continuous normal (perpendicular to surface) GRFs across a range of running speeds and slopes from accelerometer data.

METHODS

Nineteen subjects ran on a force-measuring treadmill at five slopes (0°, ±5°, ±10°) and three speeds (2.5, 3.33, 4.17 m/s) per slope with sacral- and shoe-mounted accelerometers. We then trained a recurrent neural network to predict normal GRF waveforms frame-by-frame. The predicted versus measured GRF waveforms had an average ± SD RMSE of 0.16 ± 0.04 BW and relative RMSE of 6.4 ± 1.5% across all conditions and subjects.

RESULTS

The recurrent neural network predicted continuous normal GRF waveforms across a range of running speeds and slopes with greater accuracy than neural networks implemented in previous studies. This approach may facilitate predictions of biomechanical variables outside the laboratory in near real-time and improves the accuracy of quantifying and monitoring external forces experienced by the body when running.

Foot Position Measurement during Assistive Motion for Sit-to-Stand Using a Single Inertial Sensor and Shoe-Type Force Sensors

Assistive motion for sit-to-stand causes lower back pain (LBP) among caregivers. Considering previous studies that showed that foot position adjustment could reduce lumbar load during assistive motion for sit-to-stand, quantitative monitoring of and instructions on foot position could contribute toward reducing LBP among caregivers. The present study proposes and evaluates a new method for the quantitative measurement of foot position during assistive motion for sit-to-stand using a few wearable sensors that are not limited to the measurement area. The proposed method measures quantitative foot position (anteroposterior and mediolateral distance between both feet) through a machine learning technique using features obtained from only a single inertial sensor on the trunk and shoe-type force sensors. During the experiment, the accuracy of the proposed method was investigated by comparing the obtained values with those from an optical motion capture system. The results showed that the proposed method produced only minor errors (less than 6.5% of body height) when measuring foot position during assistive motion for sit-to-stand. Furthermore, Bland–Altman plots suggested no fixed errors between the proposed method and the optical motion capture system. These results suggest that the proposed method could be utilized for measuring foot position during assistive motion for sit-to-stand.

A Comparison of Three Neural Network Approaches for Estimating Joint Angles and Moments from Inertial Measurement Units

The application of artificial intelligence techniques to wearable sensor data may facilitate accurate analysis outside of controlled laboratory settings—the holy grail for gait clinicians and sports scientists looking to bridge the lab to field divide. Using these techniques, parameters that are difficult to directly measure in-the-wild, may be predicted using surrogate lower resolution inputs. One example is the prediction of joint kinematics and kinetics based on inputs from inertial measurement unit (IMU) sensors. Despite increased research, there is a paucity of information examining the most suitable artificial neural network (ANN) for predicting gait kinematics and kinetics from IMUs. This paper compares the performance of three commonly employed ANNs used to predict gait kinematics and kinetics: multilayer perceptron (MLP); long short-term memory (LSTM); and convolutional neural networks (CNN). Overall high correlations between ground truth and predicted kinematic and kinetic data were found across all investigated ANNs. However, the optimal ANN should be based on the prediction task and the intended use-case application. For the prediction of joint angles, CNNs appear favourable, however these ANNs do not show an advantage over an MLP network for the prediction of joint moments. If real-time joint angle and joint moment prediction is desirable an LSTM network should be utilised.

Wearables and Deep Learning Classify Fall Risk From Gait in Multiple Sclerosis

Falls are a significant problem for persons with multiple sclerosis (PwMS). Yet fall prevention interventions are not often prescribed until after a fall has been reported to a healthcare provider. While still nascent, objective fall risk assessments could help in prescribing preventative interventions. To this end, retrospective fall status classification commonly serves as an intermediate step in developing prospective fall risk assessments. Previous research has identified measures of gait biomechanics that differ between PwMS who have fallen and those who have not, but these biomechanical indices have not yet been leveraged to detect PwMS who have fallen. Moreover, they require the use of laboratory-based measurement technologies, which prevent clinical deployment. Here we demonstrate that a bidirectional long short-term (BiLSTM) memory deep neural network was able to identify PwMS who have recently fallen with good performance (AUC of 0.88) based on accelerometer data recorded from two wearable sensors during a one-minute walking task. These results provide substantial improvements over machine learning models trained on spatiotemporal gait parameters (21% improvement in AUC), statistical features from the wearable sensor data (16%), and patient-reported (19%) and neurologist-administered (24%) measures in this sample. The success and simplicity (two wearable sensors, only one-minute of walking) of this approach indicates the promise of inexpensive wearable sensors for capturing fall risk in PwMS.

Biological Hip Torque Estimation using a Robotic Hip Exoskeleton

Machine learning (ML) algorithms present an opportunity to estimate joint kinetics using a limited set of mechanical sensors. These estimates could be used as a continuous reference signal for exoskeleton control, able to modulate exoskeleton assistance in real-world environments. In this study, sagittal plane biological hip torque during level ground, incline and decline walking was calculated using inverse dynamics of human subject data. Subsequently, this torque was estimated using neural network (NN) and XGBoost ML models. Model inputs consisted solely of mechanical sensor data onboard a robotic hip exoskeleton. These results were compared to a baseline method of estimating hip torque as the mean torque profile during ambulation. On average across conditions, the NN and XGBoost models estimated biological hip torque with an RMSE of 0.116±0.015 and 0.108±0.011 Nm/kg, respectively, which was significantly less than the baseline estimation that had an RMSE of 0.300±0.145 Nm/kg (p<0.05). Fitting the baseline method to ambulation mode specific data significantly reduced overall RMSE by 59.3%; however, the ML models were still significantly better than the baseline method (p<0.05). These results show that machine learning algorithms can estimate biological hip torque using only mechanical sensors onboard a hip exoskeleton better than simply using an average torque profile. This suggests that these estimation models could be suitable for modulating exoskeleton assistance. Additionally, no evidence suggested the need to train separate ML models for each ambulation mode as estimation RMSE was not significantly different across unified and separated ML models.

Modeling Muscle Synergies as a Gaussian Process: Estimating Unmeasured Muscle Excitations using a Measured Subset

Continuous observation of muscle activity could provide a comprehensive picture of the loads experienced by muscles and joints during daily life. However, a major limitation to the practical application of this approach is the need to have surface electromyography (sEMG) sensors on all involved muscles. In this work, we model the synergistic relationship between muscles as a Gaussian process enabling the inference of unmeasured muscle excitations using a subset of measured data. Specifically, we developed a model for a single subject which uses sEMG data from four leg muscles to estimate the muscle excitation time-series of six other leg muscles during level walking at a self-selected speed. The proposed technique was able to accurately estimate the held-out muscle excitation time-series of the six muscles with correlation coefficients ranging from 0.74 to 0.87 and with mean absolute error less than 3%.

A Gaussian Process Model of Muscle Synergy Functions for Estimating Unmeasured Muscle Excitations Using a Measured Subset

Estimation of muscle excitations from a reduced sensor array could greatly improve current techniques in remote patient monitoring. Such an approach could allow continuous monitoring of clinically relevant biomechanical variables that are ideal for personalizing rehabilitation. In this paper, we introduce the notion of a muscle synergy function which describes the synergistic relationship between a subset of muscles. We develop from first principles an approximation to their behavior using Gaussian process regression and demonstrate the utility of the technique for estimating the excitation time-series of leg muscles during normal walking for nine healthy subjects. Specifically, excitations for six muscles were estimated using surface electromyography (sEMG) data during a finite time interval (called the input window) from four different muscles (called the input muscles) with mean absolute error (MAE) less than 5.0% of the maximum voluntary contraction (MVC) and that accounts for 82-88% of the variance (VAF) in the true excitations. Further, these estimated excitations informed muscle activations with less than 4.0% MAE and 89-93% VAF. We also present a detailed analysis of a number of different modeling choices, including every possible combination of four-, three- and two-muscle input sets, the size and structure of the input window, and the stationarity of the Gaussian process covariance functions. Further, application specific modifications for future use are discussed. The proposed technique lays a foundation to explore the use of reduced wearable sensor arrays and muscle synergy functions for monitoring clinically relevant biomechanics during daily life.

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.