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

Enhancing accuracy and convenience of golf swing tracking with a wrist-worn single inertial sensor

In this study, we address two technical challenges to enhance golf swing trajectory accuracy using a wrist-worn inertial sensor: orientation estimation and drift error mitigation. We extrapolated consistent sensor orientation from specific address-phase signal segments and trained the estimation with a convolutional neural network. We then mitigated drift error by applying a constraint on wrist speed at the address, backswing top, and finish, and ensuring that the wrist's finish displacement aligns with a virtual circle on the 3D swing plane. To verify the proposed methods, we gathered data from twenty male right-handed golfers, including professionals and amateurs, using a driver and a 7-iron. The orientation estimation error was about 60% of the baseline, comparable to studies requiring additional sensor information or calibration poses. The drift error was halved and the single-inertial-sensor tracking performance across all swing phases was about 17 cm, on par with multimodal approaches. This study introduces a novel signal processing method for tracking rapid, wide-ranging motions, such as a golf swing, while maintaining user convenience. Our results could impact the burgeoning field of daily motion monitoring for health care, especially with the increasing prevalence of wearable devices like smartwatches.

Recurrent Neural Network Methods for Extracting Dynamic Balance Variables during Gait from a Single Inertial Measurement Unit

Monitoring dynamic balance during gait is critical for fall prevention in the elderly. The current study aimed to develop recurrent neural network models for extracting balance variables from a single inertial measurement unit (IMU) placed on the sacrum during walking. Thirteen healthy young and thirteen healthy older adults wore the IMU during walking and the ground truth of the inclination angles (IA) of the center of pressure to the center of mass vector and their rates of changes (RCIA) were measured simultaneously. The IA, RCIA, and IMU data were used to train four models (uni-LSTM, bi-LSTM, uni-GRU, and bi-GRU), with 10% of the data reserved to evaluate the model errors in terms of the root-mean-squared errors (RMSEs) and percentage relative RMSEs (rRMSEs). Independent t-tests were used for between-group comparisons. The sensitivity, specificity, and Pearson's r for the effect sizes between the model-predicted data and experimental ground truth were also obtained. The bi-GRU with the weighted MSE model was found to have the highest prediction accuracy, computational efficiency, and the best ability in identifying statistical between-group differences when compared with the ground truth, which would be the best choice for the prolonged real-life monitoring of gait balance for fall risk management in the elderly.

Saturated and asymmetric saturated control for projective synchronization of inertial neural networks with delays and discontinuous activations through matrix measure method

The projective synchronization work presented in this article is focused on a class of nonlinear discontinuous coupled inertial neural networks with mixed time-varying delays and a cluster topological structure. The synchronization problem for discontinuous coupled inertial neural networks with clustering topology is examined in consideration with the mismatched parameters and the mutual influence among various clusters. To determine the required conditions for network convergence under the influence of an extensive range of impulses, the matrix measure technique and the average impulsive intervals are employed. To illustrate the effectiveness of the theoretical findings, graphical representation of varied impulsive ranges for multiple cases are provided using numerical simulations.

Real-Time Sensor-Embedded Neural Network for Human Activity Recognition

This article introduces a novel approach to human activity recognition (HAR) by presenting a sensor that utilizes a real-time embedded neural network. The sensor incorporates a low-cost microcontroller and an inertial measurement unit (IMU), which is affixed to the subject's chest to capture their movements. Through the implementation of a convolutional neural network (CNN) on the microcontroller, the sensor is capable of detecting and predicting the wearer's activities in real-time, eliminating the need for external processing devices. The article provides a comprehensive description of the sensor and the methodology employed to achieve real-time prediction of subject behaviors. Experimental results demonstrate the accuracy and high inference performance of the proposed solution for real-time embedded activity recognition.

Automatic Body Segment and Side Recognition of an Inertial Measurement Unit Sensor during Gait

Inertial measurement unit (IMU) sensors are widely used for motion analysis in sports and rehabilitation. The attachment of IMU sensors to predefined body segments and sides (left/right) is complex, time-consuming, and error-prone. Methods for solving the IMU-2-segment (I2S) pairing work properly only for a limited range of gait speeds or require a similar sensor configuration. Our goal was to propose an algorithm that works over a wide range of gait speeds with different sensor configurations while being robust to footwear type and generalizable to pathologic gait patterns. Eight IMU sensors were attached to both feet, shanks, thighs, sacrum, and trunk, and 12 healthy subjects (training dataset) and 22 patients (test dataset) with medial compartment knee osteoarthritis walked at different speeds with/without insole. First, the mean stride time was estimated and IMU signals were scaled. Using a decision tree, the body segment was recognized, followed by the side of the lower limb sensor. The accuracy and precision of the whole algorithm were 99.7% and 99.0%, respectively, for gait speeds ranging from 0.5 to 2.2 m/s. In conclusion, the proposed algorithm was robust to gait speed and footwear type and can be widely used for different sensor configurations.

Evaluation of center of mass estimation for obese using statically equivalent serial chain

The complex structure of the human body makes its center of mass (CoM) estimation very challenging. The typically used estimation methods usually suffer from large estimation errors when applied to bodies with structural differences. Thus, a reliable estimation method is of utmost importance. In this paper, we present a detailed evaluation of a subject-specific CoM estimation technique named Statically Equivalent Serial Chain (SESC) by investigating its estimation ability over two different groups of subjects (Fit and Obese) in comparison to the segmental analysis method. For this study, we used an IMU-based motion capture system and a force platform to record the joint angles and corresponding center of pressure (CoP) values of twenty-five participants while performing a series of static postures. The root-mean-square errors (RMSE) of SESC's estimation for both groups showed close and lower mean values, whereas the segmental analysis method showed significantly larger RMSE values in comparison to SESC (p < 0.05). In addition, we used the Bland-Altman analysis to evaluate the agreement between the two techniques and the ground truth CoP, which showed the accuracy, precision, and reliability of SESC over both groups. In contrast, the segmental analysis method did not present neither accurate nor precise estimations, as our analysis revealed considerable fixed and proportional biases.

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.

Machine Learning Strategies for Low-Cost Insole-Based Prediction of Center of Gravity during Gait in Healthy Males

Whole-body center of gravity (CG) movements in relation to the center of pressure (COP) offer insights into the balance control strategies of the human body. Existing CG measurement methods using expensive measurement equipment fixed in a laboratory environment are not intended for continuous monitoring. The development of wireless sensing technology makes it possible to expand the measurement in daily life. The insole system is a wearable device that can evaluate human balance ability by measuring pressure distribution on the ground. In this study, a novel protocol (data preparation and model training) for estimating the 3-axis CG trajectory from vertical plantar pressures was proposed and its performance was evaluated. Input and target data were obtained through gait experiments conducted on 15 adult and 15 elderly males using a self-made insole prototype and optical motion capture system. One gait cycle was divided into four semantic phases. Features specified for each phase were extracted and the CG trajectory was predicted using a bi-directional long short-term memory (Bi-LSTM) network. The performance of the proposed CG prediction model was evaluated by a comparative study with four prediction models having no gait phase segmentation. The CG trajectory calculated with the optoelectronic system was used as a golden standard. The relative root mean square error of the proposed model on the 3-axis of anterior/posterior, medial/lateral, and proximal/distal showed the best prediction performance, with 2.12%, 12.97%, and 12.47%. Biomechanical analysis of two healthy male groups was conducted. A statistically significant difference between CG trajectories of the two groups was shown in the proposed model. Large CG sway of the medial/lateral axis trajectory and CG fall of the proximal/distal axis trajectory is shown in the old group. The protocol proposed in this study is a basic step to have gait analysis in daily life. It is expected to be utilized as a key element for clinical applications.

AI Prediction of Brain Signals for Human Gait Using BCI Device and FBG Based Sensorial Platform for Plantar Pressure Measurements

Artificial intelligence (AI) in developing modern solutions for biomedical problems such as the prediction of human gait for human rehabilitation is gaining ground. An attempt was made to use plantar pressure information through fiber Bragg grating (FBG) sensors mounted on an in-sole, in tandem with a brain-computer interface (BCI) device to predict brain signals corresponding to sitting, standing and walking postures of a person. Posture classification was attained with an accuracy range between 87–93% from FBG and BCI signals using machine learning models such as K-nearest neighbor (KNN), logistic regression (LR), support vector machine (SVM), and naïve Bayes (NB). These models were used to identify electrodes responding to sitting, standing and walking activities of four users from a 16 channel BCI device. Six electrode positions based on the 10–20 system for electroencephalography (EEG) were identified as the most sensitive to plantar activities and found to be consistent with clinical investigations of the sensorimotor cortex during foot movement. A prediction of brain EEG corresponding to given FBG data with lowest mean square error (MSE) values (0.065–0.109) was made with the selection of a long-short term memory (LSTM) machine learning model when compared to the recurrent neural network (RNN) and gated recurrent unit (GRU) models.

Towards Improved Inertial Navigation by Reducing Errors Using Deep Learning Methodology

Autonomous vehicles make use of an Inertial Navigation System (INS) as part of vehicular sensor fusion in many situations including GPS-denied environments such as dense urban places, multi-level parking structures, and areas with thick tree-coverage. The INS unit incorporates an Inertial Measurement Unit (IMU) to process the linear acceleration and angular velocity data to obtain orientation, position, and velocity information using mechanization equations. In this work, we describe a novel deep-learning-based methodology, using Convolutional Neural Networks (CNN), to reduce errors from MEMS IMU sensors. We develop a CNN-based approach that can learn from the responses of a particular inertial sensor while subject to inherent noise errors and provide near real-time error correction. We implement a time-division method to divide the IMU output data into small step sizes to make the IMU outputs fit the input format of the CNN. We optimize the CNN approach for higher performance and lower complexity that would allow its implementation on ultra-low power hardware such as microcontrollers. Our results show that we achieved up to 32.5% error improvement in straight-path motion and up to 38.69% error improvement in oval motion compared with the ground truth. We examined the performance of our CNN approach under various situations with IMUs of various performance grades, IMUs of the same type but different manufactured batch, and controlled, fixed, and uncontrolled vehicle motion paths.

Lower-limb Joint Torque Prediction using LSTM Neural Networks and Transfer Learning

Estimation of joint torque during movement provides important information in several settings, such as effect of athletes' training or of a medical intervention, or analysis of the remaining muscle strength in a wearer of an assistive device. The ability to estimate joint torque during daily activities using wearable sensors is increasingly relevant in such settings. In this study, lower limb joint torques during ten daily activities were predicted by long short-term memory (LSTM) neural networks and transfer learning. LSTM models were trained with muscle electromyography signals and lower limb joint angles. Hip flexion/extension, hip abduction/adduction, knee flexion/extension and ankle dorsiflexion/plantarflexion torques were predicted. The LSTM models' performance in predicting torque was investigated in both intra-subject and inter-subject scenarios. Each scenario was further divided into intra-task and inter-task tests. We observed that LSTM models could predict lower limb joint torques during various activities accurately with relatively low error (root mean square error ≤ 0.14 Nm/kg, normalized root mean square error ≤8.7%) either through a uniform model or through ten separate models in intra-subject tests. Furthermore, a transfer learning technique was adopted in the inter-task and inter-subject tests to further improve the generalizability of LSTM models by pre-training a model on multiple subjects and/or tasks and transferring the learned knowledge to a target task/subject. Particularly in the inter-subject tests, we could predict joint torques accurately in several movements after training from only a few movements from new subjects.

Deep learning approach to estimate foot pressure distribution in walking with application for a cost-effective insole system

Background

Foot pressure distribution can be used as a quantitative parameter for evaluating anatomical deformity of the foot and for diagnosing and treating pathological gait, falling, and pressure sores in diabetes. The objective of this study was to propose a deep learning model that could predict pressure distribution of the whole foot based on information obtained from a small number of pressure sensors in an insole.

Methods

Twenty young and twenty older adults walked a straight pathway at a preferred speed with a Pedar-X system in anti-skid socks. A long short-term memory (LSTM) model was used to predict foot pressure distribution. Pressure values of nine major sensors and the remaining 90 sensors in a Pedar-X system were used as input and output for the model, respectively. The performance of the proposed LSTM structure was compared with that of a traditionally used adaptive neuro-fuzzy interference system (ANFIS). A low-cost insole system consisting of a small number of pressure sensors was fabricated. A gait experiment was additionally performed with five young and five older adults, excluding subjects who were used to construct models. The Pedar-X system placed parallelly on top of the insole prototype developed in this study was in anti-skid socks. Sensor values from a low-cost insole prototype were used as input of the LSTM model. The accuracy of the model was evaluated by applying a leave-one-out cross-validation.

Results

Correlation coefficient and relative root mean square error (RMSE) of the LSTM model were 0.98 (0.92 ~ 0.99) and 7.9 ± 2.3%, respectively, higher than those of the ANFIS model. Additionally, the usefulness of the proposed LSTM model for fabricating a low-cost insole prototype with a small number of sensors was confirmed, showing a correlation coefficient of 0.63 to 0.97 and a relative RMSE of 12.7 ± 7.4%.

Conclusions

This model can be used as an algorithm to develop a low-cost portable smart insole system to monitor age-related physiological and anatomical alterations in foot. This model has the potential to evaluate clinical rehabilitation status of patients with pathological gait, falling, and various foot pathologies when more data of patients with various diseases are accumulated for training.

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.

Machine Learning Model to Estimate Net Joint Moments during Lifting Task Using Wearable Sensors: A Preliminary Study for Design of Exoskeleton Control System

Accurately measuring the lower extremities and L5/S1 moments is important since L5/S1 moments are the principal parameters that measure the risk of musculoskeletal diseases during lifting. In this study, protocol that predicts lower extremities and L5/S1 moments with an insole sensor was proposed to replace the prior methods that have spatial constraints. The protocol is hierarchically composed of a classification model and a regression model to predict joint moments. Additionally, a single LSTM model was developed to compare with proposed protocol. To optimize hyperparameters of the machine learning model and input feature, Bayesian optimization method was adopted. As a result, the proposed protocol showed a relative root mean square error (rRMSE) of 8.06~13.88% while the single LSTM showed 9.30~18.66% rRMSE. This protocol in this research is expected to be a starting point for developing a system for estimating the lower extremity and L5/S1 moment during lifting that can replace the complex prior method and adopted to workplace environments. This novel study has the potential to precisely design a feedback iterative control system of an exoskeleton for the appropriate generation of an actuator torque.

Prediction of Tool Forces in Manual Grinding Using Consumer-Grade Sensors and Machine Learning

Tool forces are a decisive parameter for manual grinding with hand-held power tools, which can be used to determine the productivity, quality of the work result, vibration exposition, and tool lifetime. One approach to tool force determination is the prediction of tool forces via measured operating parameters of a hand-held power tool. The problem is that the accuracy of tool force prediction with consumer-grade sensors remains unclear in manual grinding. Therefore, the accuracy of tool force prediction using Gaussian process regression is examined in a study for two hand-held angle grinders in four different applications in three directions using measurement data from an inertial measurement unit, a current sensor, and a voltage sensor. The prediction of the grinding normal force (rMAE = 11.44% and r = 0.84) and the grinding tangential force (rMAE = 18.21% and r = 0.82) for three tested applications, as well as the radial force for the application cutting with a cut-off wheel (rMAE = 19.67% and r = 0.80) is shown to be feasible. The prediction of the guiding force (rMAE = 87.02% and r = 0.37) for three tested applications is only possible to a limited extent. This study supports data acquisition and evaluation of hand-held power tools using consumer-grade sensors, such as an inertial measurement unit, in real-world applications, resulting in new potentials for product use and product development.

The Use of Synthetic IMU Signals in the Training of Deep Learning Models Significantly Improves the Accuracy of Joint Kinematic Predictions

Gait analysis based on inertial sensors has become an effective method of quantifying movement mechanics, such as joint kinematics and kinetics. Machine learning techniques are used to reliably predict joint mechanics directly from streams of IMU signals for various activities. These data-driven models require comprehensive and representative training datasets to be generalizable across the movement variability seen in the population at large. Bottlenecks in model development frequently occur due to the lack of sufficient training data and the significant time and resources necessary to acquire these datasets. Reliable methods to generate synthetic biomechanical training data could streamline model development and potentially improve model performance. In this study, we developed a methodology to generate synthetic kinematics and the associated predicted IMU signals using open source musculoskeletal modeling software. These synthetic data were used to train neural networks to predict three degree-of-freedom joint rotations at the hip and knee during gait either in lieu of or along with previously measured experimental gait data. The accuracy of the models’ kinematic predictions was assessed using experimentally measured IMU signals and gait kinematics. Models trained using the synthetic data out-performed models using only the experimental data in five of the six rotational degrees of freedom at the hip and knee. On average, root mean square errors in joint angle predictions were improved by 38% at the hip (synthetic data RMSE: 2.3°, measured data RMSE: 4.5°) and 11% at the knee (synthetic data RMSE: 2.9°, measured data RMSE: 3.3°), when models trained solely on synthetic data were compared to measured data. When models were trained on both measured and synthetic data, root mean square errors were reduced by 54% at the hip (measured + synthetic data RMSE: 1.9°) and 45% at the knee (measured + synthetic data RMSE: 1.7°), compared to measured data alone. These findings enable future model development for different activities of clinical significance without the burden of generating large quantities of gait lab data for model training, streamlining model development, and ultimately improving model performance.

Review of Wearable Devices and Data Collection Considerations for Connected Health

Wearable sensor technology has gradually extended its usability into a wide range of well-known applications. Wearable sensors can typically assess and quantify the wearer’s physiology and are commonly employed for human activity detection and quantified self-assessment. Wearable sensors are increasingly utilised to monitor patient health, rapidly assist with disease diagnosis, and help predict and often improve patient outcomes. Clinicians use various self-report questionnaires and well-known tests to report patient symptoms and assess their functional ability. These assessments are time consuming and costly and depend on subjective patient recall. Moreover, measurements may not accurately demonstrate the patient’s functional ability whilst at home. Wearable sensors can be used to detect and quantify specific movements in different applications. The volume of data collected by wearable sensors during long-term assessment of ambulatory movement can become immense in tuple size. This paper discusses current techniques used to track and record various human body movements, as well as techniques used to measure activity and sleep from long-term data collected by wearable technology devices.

Deep neural network approach for estimating the three-dimensional human center of mass using joint angles

Human body center of mass location plays an essential role in physical therapy, especially in investigating a subject's capability to maintain balance. However, its estimation can be a very complex, costly, and time-consuming process. To overcome the complexities and reduce the hardware cost, we proposed a deep neural network model to map the measurements of body joint angles to the 3-D center of mass position. We used an inertial measurement units-based motion-capture system (Xsens MVN Awinda) to record the joint angles and center of mass positions of 22 healthy subjects. We divided the subjects into two groups and assigned them either squat or gait tasks. Then, recorded data were merged and fed to the model to increase its generalizability. We evaluated five different input combinations to assess the effect of each input on the accuracy and generalizability of the model. The accuracy and generalizability of the models were evaluated by root-mean-square errors and comparing the differences in errors for different datasets, respectively. Root-mean-square errors ranged from 4.11 mm to 18.39 mm on both training and testing datasets for different models. Besides, adding anthropometric measurements and a Boolean parameter specifying the type of motion contributed significantly to the generalizability of the model. Also, adding unnecessary joint angles had adverse effects on the network's estimations. This study showed that by using deep neural networks, the center of mass estimations could be achieved with high accuracy, and a 17 sensors motion-capture system can be replaced with only five sensors, thus reducing the cost and complexity of the process.

IMU Data and GPS Position Information Direct Fusion Based on LSTM

In recent years, the application of deep learning to the inertial navigation field has brought new vitality to inertial navigation technology. In this study, we propose a method using long short-term memory (LSTM) to estimate position information based on inertial measurement unit (IMU) data and Global Positioning System (GPS) position information. Simulations and experiments show the practicability of the proposed method in both static and dynamic cases. In static cases, vehicle stop data are simulated or recorded. In dynamic cases, uniform rectilinear motion data are simulated or recorded. The value range of LSTM hyperparameters is explored through both static and dynamic simulations. The simulations and experiments results are compared with the strapdown inertial navigation system (SINS)/GPS integrated navigation system based on kalman filter (KF). In a simulation, the LSTM method’s computed position error Standard Deviation (STD) was 52.38% of what the SINS computed. The biggest simulation radial error estimated by the LSTM method was 0.57 m. In experiments, the LSTM method computed a position error STD of 23.08% using only SINSs. The biggest experimental radial error the LSTM method estimated was 1.31 m. The position estimated by the LSTM fusion method has no cumulative divergence error compared to SINS (computed). All in all, the trained LSTM is a dependable fusion method for combining IMU data and GPS position information to estimate position.

Estimation of Human Center of Mass Position through the Inertial Sensors-Based Methods in Postural Tasks: An Accuracy Evaluation

The estimation of the body’s center of mass (CoM) trajectory is typically obtained using force platforms, or optoelectronic systems (OS), bounding the assessment inside a laboratory setting. The use of magneto-inertial measurement units (MIMUs) allows for more ecological evaluations, and previous studies proposed methods based on either a single sensor or a sensors’ network. In this study, we compared the accuracy of two methods based on MIMUs. Body CoM was estimated during six postural tasks performed by 15 healthy subjects, using data collected by a single sensor on the pelvis (Strapdown Integration Method, SDI), and seven sensors on the pelvis and lower limbs (Biomechanical Model, BM). The accuracy of the two methods was compared in terms of RMSE and estimation of posturographic parameters, using an OS as reference. The RMSE of the SDI was lower in tasks with little or no oscillations, while the BM outperformed in tasks with greater CoM displacement. Moreover, higher correlation coefficients were obtained between the posturographic parameters obtained with the BM and the OS. Our findings showed that the estimation of CoM displacement based on MIMU was reasonably accurate, and the use of the inertial sensors network methods should be preferred to estimate the kinematic parameters.

Multiple Inertial Measurement Unit Combination and Location for Center of Pressure Prediction in Gait

Center of pressure (COP) during a gait cycle indicates crucial information with regard to fall risk such as balance capacity. The drawbacks of conventional research instruments include inconvenient use during activities of daily living and expensive costs. The present study illustrates the promising fall-relevant information predicted by acceleration and angular velocity data from different placement sensors with machine learning techniques. This approach is inspired by the emerging machine learning technique, specifically the long short-term memory (LSTM), which is often used in time series data and aims to decrease the burden of the user while using the novel wearable technology. The Jaccard similarity coefficient, which implies the consistency of profile alignment between prediction and real situation, achieved 94% accuracy in the walking direction. Furthermore, the number of sensors used and the placement influenced the feasibility of an application. The outcome revealed that the accuracy could exceed 90% with only one sensor placed on the foot in the walking direction, and the toe would be the best location for sensor placement. To examine the performance of machine learning, the current study employed two parameters from different perspectives. One is a commonly used parameter, which represented the error, and the other investigated the similarity between the prediction and ground truth. From a similarity perspective, the parameter can be used as a metric to assess the consistency of profile alignment.

Acceleration Magnitude at Impact Following Loss of Balance Can Be Estimated Using Deep Learning Model

Pre-impact fall detection can detect a fall before a body segment hits the ground. When it is integrated with a protective system, it can directly prevent an injury due to hitting the ground. An impact acceleration peak magnitude is one of key measurement factors that can affect the severity of an injury. It can be used as a design parameter for wearable protective devices to prevent injuries. In our study, a novel method is proposed to predict an impact acceleration magnitude after loss of balance using a single inertial measurement unit (IMU) sensor and a sequential-based deep learning model. Twenty-four healthy participants participated in this study for fall experiments. Each participant worn a single IMU sensor on the waist to collect tri-axial accelerometer and angular velocity data. A deep learning method, bi-directional long short-term memory (LSTM) regression, is applied to predict a fall's impact acceleration magnitude prior to fall impact (a fall in five directions). To improve prediction performance, a data augmentation technique with increment of dataset is applied. Our proposed model showed a mean absolute percentage error (MAPE) of 6.69 ± 0.33% with r value of 0.93 when all three different types of data augmentation techniques are applied. Additionally, there was a significant reduction of MAPE by 45.2% when the number of training datasets was increased by 4-fold. These results show that impact acceleration magnitude can be used as an activation parameter for fall prevention such as in a wearable airbag system by optimizing deployment process to minimize fall injury in real time.

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).