Next Article in Journal
The Optical Parameter Optimization for Brain Implant Alzheimer Sensor Using Phototherapy Angle and Wavelength Simulation (PAWS) Methodology
Previous Article in Journal
Adaptive Sharding for UAV Networks: A Deep Reinforcement Learning Approach to Blockchain Optimization
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Sensors
Volume 24
Issue 22
10.3390/s24227280
sensors-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Systematic Review

The Use of Wearable Sensors and Machine Learning Methods to Estimate Biomechanical Characteristics During Standing Posture or Locomotion: A Systematic Review

by
Isabelle J. Museck
1,*,
Daniel L. Brinton
2 and
Jesse C. Dean
3,4,*
1
Department of Health Sciences and Research, Medical University of South Carolina, Charleston, SC 29425, USA
2
Department of Healthcare Leadership and Management, Medical University of South Carolina, Charleston, SC 29425, USA
3
Department of Rehabilitation Sciences, Medical University of South Carolina, Charleston, SC 29425, USA
4
Ralph H. Johnson VA Medical Center, Charleston, SC 29401, USA
*
Authors to whom correspondence should be addressed.
Sensors 2024, 24(22), 7280; https://doi.org/10.3390/s24227280
Submission received: 11 September 2024 / Revised: 25 October 2024 / Accepted: 1 November 2024 / Published: 14 November 2024
(This article belongs to the Section Wearables)
Browse Figures
Review Reports Versions Notes

Abstract

:
Balance deficits are present in a variety of clinical populations and can negatively impact quality of life. The integration of wearable sensors and machine learning technology (ML) provides unique opportunities to quantify biomechanical characteristics related to balance outside of a laboratory setting. This article provides a general overview of recent developments in using wearable sensors and ML to estimate or predict biomechanical characteristics such as center of pressure (CoP) and center of mass (CoM) motion. This systematic review was conducted according to PRISMA guidelines. Databases including Scopus, PubMed, CINHAL, Trip PRO, Cochrane, and Otseeker databases were searched for publications on the use of wearable sensors combined with ML to predict biomechanical characteristics. Fourteen publications met the inclusion criteria and were included in this review. From each publication, information on study characteristics, testing conditions, ML models applied, estimated biomechanical characteristics, and sensor positions were extracted. Additionally, the study type, level of evidence, and Downs and Black scale score were reported to evaluate methodological quality and bias. Most studies tested subjects during walking and utilized some type of neural network (NN) ML model to estimate biomechanical characteristics. Many of the studies focused on minimizing the necessary number of sensors and placed them on areas near or below the waist. Nearly all studies reporting RMSE and correlation coefficients had values <15% and >0.85, respectively, indicating strong ML model estimation accuracy. Overall, this review can help guide the future development of ML algorithms and wearable sensor technologies to estimate postural mechanics.

1. Introduction

Balance deficits are common in many clinical populations, specifically among individuals with lower extremity musculoskeletal injuries [1], stroke [2], degenerative diseases [3], and advanced age [4]. An estimated 13% of adults between ages 65 and 69 self-report imbalance and this proportion increases to 46% in those aged 85 and older [5]. Additionally, balance issues are often precursors of falls [6], which are the leading cause of injury-related mortality among older adults in the United States [7]. Deteriorating control of balance can greatly impact functional mobility [8], quality of life, and independence [9].
Biomechanical measures of gait and standing posture can assess a patient’s balance function and provide insight into potential rehabilitation goals or treatments [10]. The current gold standard systems used to evaluate biomechanical parameters of gait and postural balance, such as force platforms and motion capture systems, are complex, expensive, and often require highly trained operators, preventing translation outside the laboratory and clinic [11,12]. These systems often quantify characteristics of body motion based on the trajectory of the center of pressure (CoP) or center of mass (CoM) [12]. These biomechanical characteristics can provide insight into the underlying balance control during either standing or walking and can be used to assess the effects of rehabilitation interventions [13,14].
Recently, substantial work has focused on the development of portable technology that allows the measurement of biomechanical characteristics outside of a laboratory setting. Wearable technologies offer a low-cost, portable way to track biomechanical characteristics in real-time [12], with typical wearable sensors such as inertial measurement units (IMUs) including accelerometer, gyroscope, and magnetometer components. Several systematic reviews have addressed the use of wearable devices to track biomechanical characteristics and estimate fall risk for older adults. A systematic review from 2017 investigated the use of mobile phone technology to assess balance and fall risk [15], while another from 2018 addressed different novel sensing technologies used for fall risk assessment in older adults [16]. More recently, a review from 2019 focused on providing a detailed overview of the common applications of wearable sensors to assess postural balance [12]. The use of wearable sensors to assess balance, however, is an area of research that is continually evolving. While existing reviews focus on using data collected directly from wearable sensors to assess balance or fall risk, there is a noticeable gap in research evaluating the use of wearables to estimate or predict balance characteristics.
The multi-dimensional, multi-correlated, and patient-specific nature of the data collected from postural and gait analysis can challenge analysis and comprehension. One possible approach to overcoming this challenge may be to use machine learning (ML) and artificial intelligence (AI) methods to process data to predict or estimate biomechanical characteristics. Prediction involves forecasting future or unseen data values from previous training data while estimations approximate current data or parameters based on historical data. In the healthcare sector, ML has rapidly gained attention due to its demonstrated potential to solve problems and perform certain medical tasks such as diagnosis and detection more efficiently and effectively than humans [17]. However, the use of ML models to assess human gait and postural balance is a relatively new area of research [18]. While such ML applications were briefly discussed in the 2019 review of wearable sensors [12], there is a lack of an updated and comprehensive overview addressing the integration of wearable sensors and machine learning technology to estimate or predict continuous biomechanical characteristics such as CoP and CoM motion.
The purpose of this research is to provide an overview of recent work that has incorporated wearable sensors with machine learning technology to estimate or predict biomechanical characteristics (CoP and CoM) while standing or walking. We thus extend upon the previous systematic reviews published in the last few years that focused on more traditional approaches to IMU-based balance assessment. This work provides a detailed review of emerging research and technologies that will likely be a component of future personalized, adaptive, and portable rehabilitation therapies for those with balance, posture, and gait abnormalities. Through the selection of high-quality papers that combine machine learning technology and wearable sensors to quantitatively estimate biomechanical characteristics, this study aims to (1) investigate optimal machine learning models for estimating CoP and CoM from wearable sensor data inputs, (2) identify common locations for sensor placement and testing conditions, (3) identify common ML model performance metrics across studies, and, finally, (4) suggest future directions of design and implementation of wearable sensors and ML models to accurately estimate and predict postural mechanics. This review is organized into five sections: Section 1 provides the background and rationale for the study; Section 2 outlines the methodology for searching, selecting articles for inclusion, extracting data, and assessing quality; Section 3 presents the results and characteristics of the included articles; Section 4 discusses these findings and the study limitations; and Section 5 concludes with the implications and future directions of this research.

2. Materials and Methods

2.1. Search Strategy

A literature search was conducted using the following six databases: Scopus, PubMed, CINAHL, Trip PRO, Cochrane, and OTseeker. Articles including a variation of the following terms were included: “artificial intelligence algorithms”, “wearable sensors”, and “biomechanical characteristics”. Title and abstract keywords and their synonyms were converted from the different database search syntaxes accordingly. The interest in ML for healthcare has increased rapidly over the last 10 years. Following 2010, the number of clinical practice and clinical research-related publications with machine learning increased exponentially [19]. To evaluate recent trends, only articles published from January 2010 to February 2024 were included in this review. This search includes English-language, peer-reviewed journal articles only. The computerized database search was performed using the following query in Table 1.

2.2. Study Selection

The method of selecting relevant articles consisted of the following process given by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [20]. Details regarding the necessary reporting items can be found in the PRISMA 2020 checklist in the Supplementary Materials. Each database was last consulted on 2 February 2024. Using the EndNote reference manager software version 21, duplicates were removed, and the remaining articles were hand-screened by their titles and abstracts. To reduce the risk of bias, a random selection of 50 articles from the original search (with duplicates removed) was sent to an outside reviewer for additional title and abstract screening. Following title and abstract screening, articles were read fully and filtered out by the primary reviewer according to the following inclusion and exclusion criteria.
Inclusion Criteria:
  • Only peer-reviewed, web-available journal articles.
  • Articles published in the period from January 2010 to February 2024.
  • Studies that use wearable sensor data as inputs for model prediction.
  • Articles using machine learning to estimate or predict CoP, CoM, or a variant of these two balance-related biomechanical characteristics.
Exclusion Criteria:
  • Reviews, conference papers, magazines, or book chapter papers.
  • Studies in languages other than English.
  • Publications that studied robotic, exoskeleton-based systems, or non-human subjects.
  • Studies that used non-wearable sensors-based systems, such as cameras and radar to make predictions.
  • Studies that used ML and wearable sensors solely to detect instability, falls, freezing of gait, or assess fall risk.
  • Studies that used ML and wearable sensors to classify different postural behaviors or postural statuses.
  • Studies that used ML and wearable sensors to estimate PT labels or clinical rating scales of balance performance.
  • Studies that used ML and wearable sensors solely to predict joint angles, torques, and ground reaction forces (GRFs).

2.3. Data Extraction

After confirming with a collaborator, the data items needed for meaningful synthesis were collected from each report. Study objectives, participant health status, age, sample size, and level of evidence were all extracted from the selected articles. Additionally, to facilitate analysis and investigate the current state of using machine learning models and wearable sensors to estimate biomechanical characteristics, the following information was extracted from the selected articles: (1) Machine learning model-related information: including the type of model used, features extracted, and training method. (2) Sensor-related information: including the sensor type, number, and locations during signal acquisition. (3) Testing condition information: data were acquired under various testing conditions (walking, static standing, and running). (4) Duration of collected biomechanical data, as in the duration of the trials for each testing condition. (5) Biomechanical characteristic of interest: different studies focused on the estimation and prediction of several different biochemical parameters including the center of pressure (CoP), the center of mass (CoM), the center of mass height (CoMH), the center of pressure–center of mass inclination angle (CoP-CoM IA), and the strike index. The strike index is defined as the location of the center of pressure along the long axis of the foot as a percentage of the total foot length at the initial foot strike [21]. Different outcome-related information such as the error ratio (ER), root-mean-square errors (RMSEs), normalized root-mean-square errors (NRMSEs), Jaccard index [22], and R-squared values were used to report the performance of the machine learning model predictions and estimations. The primary outcomes sought from each study were these performance metrics demonstrating the ML model’s ability to predict characteristics like CoP or CoM displacement, velocity, location, height, or angle. The extracted data were tabulated to help compare study properties and results. Missing summary or unclear data were analyzed by the reviewer and reasonable inferences were made based on accompanying information.

2.4. Quality Assessment

Each article was screened and scored using the Downs and Black Quality checklist to assess overall methodological quality, internal and external validity, and power of the study [23]. The Downs and Black checklist consists of 27 items distributed between 5 sub-scales: reporting, external validity, bias, confounding, and power. The articles were scored Yes = 1, No = 0, and Unclear/NA = 0 for each item on the checklist. Investigating the differences in sample and study characteristics such as subject age, health condition, type of ML model applied, test condition, and estimated biomechanical characteristic allowed for exploration of possible causes of heterogeneity among the study results. Studies with similar samples, characteristics, and outcome measures were grouped for comparison and analysis. The level of evidence of each study was also investigated to assess confidence and certainty in the results. Each study design was categorized based on a recently developed pyramid for classifying AI scientific evidence presented by Bellini and colleagues shown in Figure 1 [24].

3. Results

A detailed flow diagram of the search results and screening strategy is presented in Figure 2. The initial search resulted in 192 articles for possible inclusion in the present review. After the removal of duplicated studies, 131 articles remained for title and abstract screening. The initial screening of titles and abstracts removed 96 studies, and the remaining 35 articles were reviewed in their full-text versions to assess their inclusion in the review. The full-text screening removed 21 studies due to the previously stated exclusion criteria, which involved (i) studies that used IMUs and ML to classify postural status (such as sitting, standing, and walking) [25,26], (ii) studies that focused on classifying postural behaviors [27,28,29,30,31] or predicting joint kinetics [32], (iii) studies that assessed fall risk or detected falls [33,34], (iv) studies that estimated clinical scales or PT ratings [35,36,37,38] (v) studies that used non-human subjects [39,40,41], (vi) and studies that were not peer-reviewed [42,43,44,45]. The rigorous review process resulted in 14 quality articles that were included in this systematic review.
Table 2 summarizes the data extracted from the 14 articles, including the author and year of publication, characteristics of the study population, testing conditions, machine learning models applied, estimated biomechanical characteristics, and sensor-related information. The study type, level of evidence, and the Downs and Black scale score out of 27 points are also reported in Table 2.

3.1. Study Design and Methodological Quality

The distribution of methodological quality and bias assessment scores based on the Downs and Black checklist are presented in Table 2. The reviewed articles had an average Downs and Black score of 15 with scores ranging between 11 and 21, and over half of the scores ranging between 13 and 15 points. All the studies, with the exclusion of one [48], were considered internal validation studies according to the AI scientific evidence pyramid [24] and were classified as level III evidence using Sackett’s criteria [58].

3.2. Outcome Measurements

There are several different methods by which the studies in this review reported results and outcomes of ML model predictions. The results from each study quantifying the ML model’s accuracy in predicting the biomechanical characteristic of interest (CoP, CoM, CoP-CoM IA, etc.) are represented in Table 3. Current ML-based IMU literature widely uses the root mean square error (RMSE) to evaluate the prediction accuracy of the estimated gait variables [55]. All studies, except three [47,48,50], reported prediction outcomes in terms of RMSE and/or normalized root-mean-square error (NRMSE). RMSE is reported in units of the response variable, or biomechanical characteristic in this case, and is calculated by taking the square root of the average of the squared differences between the true or actual values and the predicted values. The NRMSE is simply the RMSE divided by the range of values and is reported as a percentage. Lower RMSE and NRMSE values indicate higher model prediction accuracy. Both measurements are useful for comparing different models with one another. Many of the studies (64%) used a measure related to correlation coefficient such as R-squared or Spearman’s Rho [47] to evaluate a model’s ability to estimate gait or standing biomechanical characteristics. Correlation coefficients are useful in determining whether a model is objectively a good fit for the data. Coefficients close to unity (1.0) indicate strong relationships and lower model errors. One study [50] reported outcomes as an error ratio (ER) percentage of the estimated CoP and true CoP, while another study [54] reported model prediction results in terms of the Jaccard Index, which is a measure of similarity between two datasets determined by calculating the percentage of the size of the intersection between ground truth and predicted CoP (the number in both sets) divided by the size of union between the two (the number in either set). Although the performance of the models in each of the studies can be evaluated individually, it is difficult to compare the performance of these models across studies due to the differences in various parameters such as inputs, outputs, validation, and model training data.

3.3. Machine Learning Models

The 14 reviewed studies applied a variety of ML models to estimate biomechanical characteristics from collected wearable sensor data. Many of the studies applied and tested more than one ML model. A large majority of the machine learning applications (about 77%) were neural network (NN) models. Neural networks are a type of ML model that mimics biological neuroscience and uses layers of interconnected neurons to learn from examples and make decisions, recognize patterns, or solve problems [59]. More specifically, of the 22 applications of machine learning, 7 studies applied a type of long short-term memory (LSTM) modeling. LSTM models are a type of neural network that uses memory cells to learn long-term dependencies in sequential data. This structure makes such models suitable for time series predictions and estimating changing parameters like CoM and CoP [56]. Each model’s frequency of use in the reviewed studies is summarized in Figure 3.

3.4. Wearable Sensor Properties

Two types of sensors were primarily used in these studies: inertial measurement units (IMUs) and triaxial accelerometers. Almost 80% of the studies used IMU sensors, which are typically composed of accelerometer, magnetometer, and gyroscope components that capture motion signals. There was one study [46] that used IMUs in combination with force-sensing resistors (FSRs) as input for ML predictions. In terms of sensor location, the lower body including the feet, shanks, thighs, and hip/lower back region were the most common locations to which sensors were attached. There is a variety in the number of sensors used to make predictions among the studies in this review; however, many studies had the stated goal of minimizing the number of sensors. Figure 4 indicates the locations of sensor placement and Figure 5 summarizes the characteristics of the wearables used in conjunction with the ML models to predict biomechanical characteristics in the reviewed articles.

3.5. Testing Activity

Walking (60%) was the primary testing activity used to explore the use of ML models and wearable sensors to predict biomechanical characteristics. Standing balance (33%) was another common condition tested. This was often not just simple, static, bipedal standing but rather standing on one foot or in tandem [47], standing while leaning in the AP/ML directions [50], or standing while receiving external perturbations [48,52]. Only one article [21] focused on the estimation of biomechanical characteristics during running. The different test activities may cause differences within the results from these studies and can make it difficult to make direct comparisons as prediction accuracy may be affected by varying activities like standing, walking, and running.

3.6. Predicted Biomechanical Characteristic of Posture

Half of the studies (50%) in this review focused on using ML models and wearable sensors to predict or estimate a subject’s center of pressure (CoP). Five of the studies focused on the prediction of certain center of mass measurements (CoM) such as CoM motion, CoM height, or CoM acceleration, and two of the studies involved the incorporation of both CoP and CoM characteristics to predict the CoP-CoM inclination angle (CoP-CoM IA). The prediction of different biomechanical characteristics is an additional cause for variations between the results of these studies.

4. Discussion

This systemic review examined literature that estimated biomechanical characteristics such as CoP and CoM using machine learning, with a focus on the integration of sensor technology and ML models. There were a variety of patient populations, testing protocols, and characteristics examined among the studies included in this review. Although multiple different types of ML models and sensor configurations were used to make predictions across studies, nearly all the authors reported favorable ML model prediction and estimation outcomes.

4.1. Participant Characteristics

Almost all the papers included in this review involved healthy individuals in their mid-twenties, while few included participants with a neurodegenerative condition such as stroke or a spinal cord injury. Many previous studies involving the initial investigation of the integration of ML models and wearable sensors similarly used healthy or able-bodied subjects in their studies [60]. It is common to begin ML model exploration with healthy control subjects and then move into more complex systems such as patients with different neurological conditions. In a recent study investigating the use of ML models for gait activity classification, Slemenšek and colleagues first focused on identifying the best-performing ML algorithm for activity classification of a single 25-year-old male subject and then expanded to a larger dataset of five subjects, with a final follow-up proof of concept on patients with neurological pathologies [61]. The goal of much of this research is to ultimately develop models to help those with gait or balance impairments; therefore, the articles in this review that tested their ML models on both healthy controls and pathological subjects [46,49,57] are of additional value.

4.2. Testing Parameters and Sensor Properties

The task performed by participants can have a considerable effect on the accuracy of ML model predictions, as can the number and location of wearable sensors. Sensor number and placement varied from study to study; however, the most frequent sensor placement locations across all the studies were the feet, shanks, thighs, and hip/lower back region. This is consistent with the most common sensor placement locations from a recent systematic review of wearable sensor-based technologies for fall risk assessment [62]. Vincent et al. demonstrated that the location (upper trunk, waist, lower thigh) of an accelerometer influenced the accuracy of CoP estimation during standing, regardless of the different ML models applied [50]. Of the three separate locations they tested in the study, it was determined that the waist was the optimal position for sensor placement to estimate CoP while standing. One of the studies in this review tested model performance with multiple different sensor locations and found that when using one IMU for estimation of CoP during walking, accuracy levels ranked from best to worst were back, right thigh, left thigh, left shank, right shank, right foot, and left foot [49]. Similarly, Howcroft et al. determined that sensor data collected from areas near the center of mass (i.e., lumbar and waist) were most sensitive and performed the best for identifying fall risk compared with other locations [63]. A separate study [61] determined sensor placement below the knee joint, anterior of the shin bone was superior to more proximal placement in terms of differentiating different gait activities rather than estimating a continuous parameter like CoP or CoM. The more noticeable differences in lower leg movements during different gait activities likely contributed to making these classifications more distinguishable. Overall, it seems that sensors placed on the lower half of the body perform the best in predicting CoM and CoP, and placement near the sacrum or waist is particularly sensitive to postural changes, making for more accurate estimations of these balance characteristics.
Across the studies, participants performed a variety of different tasks during data collection such as standing still, standing on one foot, walking, and even running. The chosen placement location of the sensors is likely largely dependent on the task performed during data collection. For example, it is reasonable to assume that assessments of standing balance would benefit little from IMU sensor data collected from the feet, as there is little to no foot movement during static standing. On the other hand, sensor data collected from the feet during walking could be very valuable input for ML model predictions.
Similar to the review from Jiao and colleagues on automatic post-stroke classification systems [64], the studies in this analysis used different numbers of sensors to train ML model predictions. Although, in theory, more sensors lead to more data and could lead to more accurate ML predictions, it was more common for the articles in this review to investigate the use of 3 or fewer wearable sensors with promising ML model prediction accuracy. This is in line with current literature that has demonstrated over 90% accuracy in gait classification and gait phase estimation using a single IMU sensor and ML models [65]. One of the studies included in this review [49] tested their ML models with differing quantities of IMU sensors and found that the number of IMUs can be reduced to five without deterioration in the model performance. Although fewer sensors are ideal as less equipment is more easily translated to real-world settings, researchers must be careful to select the proper number of sensors to ensure their desired prediction accuracy.

4.3. Limitations

Small sample sizes contribute to a risk of inaccurate results for many of the reviewed studies, which often recruited fewer than 25 participants for ML model development and even fewer for those that included stroke patients. One study reported data from a single participant [48]. Although some studies collected multiple trials from each participant to expand the data set, this way of model training may have caused overestimations of classification performance due to decreased variation in the training and test data [64]. Additionally, the low number of both healthy and pathological participant observations may have affected the ML model development and its ability to generalize to new data.
Differences in analysis methodology also complicate comparisons between the identified studies. The most common statistical measures for evaluating the performance of regression are Mean Absolute Error (MAE), Mean Squared Error (MSE), root-mean-square error (RMSE), R2, and adjusted R2 [66]. Some studies only reported one of these outcomes [47,48] and others included other metrics of reporting accuracy such as the Jaccard Index and Error Ratio [50,54], preventing direct comparison of model performance among studies. Standardization among reporting ML model performance would allow models from different studies to be compared more accurately. However, the varying goals of these models may prevent agreement on the most appropriate metric. Other factors such as differences in training data, validation processes, sample sizes, and hyperparameters can also affect the outcome metrics across the models in each study. Thus, future models could include a standard for training data set size and percentage of validation and test data to allow for consistency and valid comparison amongst ML model results. Nearly all of the studies that reported correlation coefficients and NRMSE, however, had values > 0.85 and <15%, respectively, indicating that most of the models across all the studies in this review demonstrated strong performance in estimating CoP and CoM. Liu and colleagues proposed that a machine learning model is highly accurate if the correlation coefficient is greater than 0.9 and the NRMSE is 15% or less [67]. The required model performance will depend on its application within the field. For example, a model used to detect cancer from MRI images will likely require higher prediction accuracies for clinical adoption than a model used to predict time in surgery for orthopedic patients. Although these results are promising and valuable for guiding the future development of ML models to predict CoP or CoM, it is important to consider internal and external validation methods, the size and diversity of the training data set, and the generalizability of these results to truly assess their efficacy and performance in a clinical setting and as a rehabilitation intervention. Further research is necessary for all models tested in these studies to properly demonstrate efficacy with larger sample sizes and more diverse populations.
This systematic review provides an overview of the recent literature on the development of ML technology to predict CoP and CoM using wearable sensors. Future research could involve a more detailed data extraction process from the ML models included in each study to allow for an even deeper understanding of the model performance and a stronger comparison of the results. Extracting details regarding the training data, test data, validation process, and model hyperparameters would allow for a more detailed and comprehensive review of ML model development and performance among these studies.

5. Conclusions

The results from this review indicate that ML models can accurately estimate CoP and CoM data from wearable sensor data. Despite the challenge of identifying a single optimal model due to the diverse characteristics of the studies reviewed, neural networks were the most commonly used model and have demonstrated highly accurate predictions. Most studies focused on estimating biomechanical characteristics during walking, with sensors typically placed on the feet, hips, and lower limbs. Model performance metrics were often reported as correlation coefficients or RMSE values.
Although a variety of models, testing parameters, sensors, and validation methods were used among the studies in this review, this information can be used to guide new studies in the development of optimal methods for designing and testing ML models that make estimations from wearable sensor data. Future research may focus on a more detailed analysis of machine learning model parameters and the development of standardized methods for stronger comparisons of model performance across studies. Moving forward, researchers can aim to enhance the design of these models by creating systems capable of making accurate predictions across various daily activities. Additionally, reducing the number of sensors required for these predictions will make the system more portable and user-friendly. Overall, the results of this review will help direct the development of novel methods, devices, and models for balance rehabilitation.

Supplementary Materials

The following supporting information can be downloaded at https://www.mdpi.com/article/10.3390/s24227280/s1, Table S1: PRISMA 2020 Checklist.

Author Contributions

Conceptualization, I.J.M.; methodology, I.J.M. and D.L.B.; validation, I.J.M. and J.C.D.; formal analysis, I.J.M. and J.C.D.; investigation, I.J.M.; resources; I.J.M. and D.L.B.; data curation, I.J.M.; writing—original draft preparation, I.J.M.; writing—review and editing, D.L.B., J.C.D. and I.J.M.; visualization, I.J.M.; supervision, D.L.B. and J.C.D.; project administration, J.C.D.; funding acquisition, J.C.D. All authors have read and agreed to the published version of the manuscript.

Funding

The research and APC were funded by the National Science Foundation grant number 2242812.

Acknowledgments

Writing support was received from R.J. Lambert, at the Medical University of South Carolina Center for Academic Excellence & Writing Center.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Wang, S.; Hu, X.; Rivera-Lillo, G. Editorial: The effect of musculoskeletal conditions on balance control. Front. Rehabil. Sci. 2023, 4, 1288322. [Google Scholar] [CrossRef] [PubMed]
  2. Tyson, S.F.; Hanley, M.; Chillala, J.; Selley, A.; Tallis, R.C. Balance disability after stroke. Phys. Ther. 2006, 86, 30–38. [Google Scholar] [CrossRef]
  3. Lamptey, R.N.L.; Chaulagain, B.; Trivedi, R.; Gothwal, A.; Layek, B.; Singh, J. A Review of the Common Neurodegenerative Disorders: Current Therapeutic Approaches and the Potential Role of Nanotherapeutics. Int. J. Mol. Sci. 2022, 23, 1851. [Google Scholar] [CrossRef]
  4. Osoba, M.Y.; Rao, A.K.; Agrawal, S.K.; Lalwani, A.K. Balance and gait in the elderly: A contemporary review. Laryngoscope Investig. Otolaryngol. 2019, 4, 143–153. [Google Scholar] [CrossRef] [PubMed]
  5. Gerson, L.W.; Jarjoura, D.; McCORD, G. Risk of imbalance in elderly people with impaired hearing or vision. Age Ageing 1989, 18, 31–34. [Google Scholar] [CrossRef] [PubMed]
  6. Sudarsky, L. Gait disorders: Prevalence, morbidity, and etiology. Adv. Neurol. 2001, 87, 111–117. [Google Scholar] [PubMed]
  7. Lohman, M.C.; Sonnega, A.J.; Nicklett, E.J.; Estenson, L.; Leggett, A.N. Comparing Estimates of Fall-Related Mortality Incidence Among Older Adults in the United States. J. Gerontol. A Biol. Sci. Med. Sci. 2019, 74, 1468–1474. [Google Scholar] [CrossRef]
  8. Wesselhoff, S.; Hanke, T.A.; Evans, C.C. Community mobility after stroke: A systematic review. Top. Stroke Rehabil. 2018, 25, 224–238. [Google Scholar] [CrossRef]
  9. Mahlknecht, P.; Kiechl, S.; Bloem, B.R.; Willeit, J.; Scherfler, C.; Gasperi, A.; Rungger, G.; Poewe, W.; Seppi, K. Prevalence and burden of gait disorders in elderly men and women aged 60–97 years: A population-based study. PLoS ONE 2013, 8, e69627. [Google Scholar] [CrossRef]
  10. Prakash, C.; Kumar, R.; Mittal, N. Recent developments in human gait research: Parameters, approaches, applications, machine learning techniques, datasets and challenges. Artif. Intell. Rev. 2018, 49, 1–40. [Google Scholar] [CrossRef]
  11. Das, R.; Paul, S.; Mourya, G.K.; Kumar, N.; Hussain, M. Recent Trends and Practices Toward Assessment and Rehabilitation of Neurodegenerative Disorders: Insights from Human Gait. Front. Neurosci. 2022, 16, 859298. [Google Scholar] [CrossRef]
  12. Ghislieri, M.; Gastaldi, L.; Pastorelli, S.; Tadano, S.; Agostini, V. Wearable Inertial Sensors to Assess Standing Balance: A Systematic Review. Sensors 2019, 19, 4075. [Google Scholar] [CrossRef]
  13. Labrozzi, G.C.; Warner, H.; Makowski, N.S.; Audu, M.L.; Triolo, R.J. Center of Mass Estimation for Impaired Gait Assessment Using Inertial Measurement Units. IEEE Trans. Neural Syst. Rehabil. Eng. 2024, 32, 12–22. [Google Scholar] [CrossRef]
  14. Quijoux, F.; Nicolaï, A.; Chairi, I.; Bargiotas, I.; Ricard, D.; Yelnik, A.; Oudre, L.; Bertin-Hugault, F.; Vidal, P.; Vayatis, N.; et al. A review of center of pressure (COP) variables to quantify standing balance in elderly people: Algorithms and open-access code. Physiol. Rep. 2021, 9, e15067. [Google Scholar] [CrossRef]
  15. Roeing, K.L.; Hsieh, K.L.; Sosnoff, J.J. A systematic review of balance and fall risk assessments with mobile phone technology. Arch. Gerontol. Geriatr. 2017, 73, 222–226. [Google Scholar] [CrossRef]
  16. Sun, R.; Sosnoff, J.J. Novel sensing technology in fall risk assessment in older adults: A systematic review. BMC Geriatr. 2018, 18, 14. [Google Scholar] [CrossRef]
  17. Esteva, A.; Kuprel, B.; Novoa, R.A.; Ko, J.; Swetter, S.M.; Blau, H.M.; Thrun, S. Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017, 542, 115–118. [Google Scholar] [CrossRef]
  18. Lattanzi, E.; Freschi, V. Evaluation of human standing balance using wearable inertial sensors: A machine learning approach. Eng. Appl. Artif. Intell. 2020, 94, 103812. [Google Scholar] [CrossRef]
  19. Weissler, E.H.; Naumann, T.; Andersson, T.; Ranganath, R.; Elemento, O.; Luo, Y.; Freitag, D.F.; Benoit, J.; Hughes, M.C.; Khan, F.; et al. The role of machine learning in clinical research: Transforming the future of evidence generation. Trials 2021, 22, 537. [Google Scholar] [CrossRef]
  20. Page, M.J.; McKenzie, J.E.; Bossuyt, P.M.; Boutron, I.; Hoffmann, T.C.; Mulrow, C.D.; Shamseer, L.; Tetzlaff, J.M.; Akl, E.A.; Brennan, S.E.; et al. The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ 2021, 372, n71. [Google Scholar] [CrossRef]
  21. Tan, T.; Strout, Z.A.; Cheung, R.T.; Shull, P.B. Strike index estimation using a convolutional neural network with a single, shoe-mounted inertial sensor. J. Biomech. 2022, 139, 111145. [Google Scholar] [CrossRef]
  22. Jaccard, P. Etude de la distribution florale dans une portion des Alpes et du Jura. Bull. Soc. Vaudoise Sci. Nat. 1901, 37, 547–579. [Google Scholar]
  23. Downs, S.H.; Black, N. The feasibility of creating a checklist for the assessment of the methodological quality both of randomised and non-randomised studies of health care interventions. J. Epidemiol. Community Health 1998, 52, 377–384. [Google Scholar] [CrossRef]
  24. Bellini, V.; Coccolini, F.; Forfori, F.; Bignami, E. The artificial intelligence evidence-based medicine pyramid. World J. Crit. Care Med. 2023, 12, 89–91. [Google Scholar] [CrossRef]
  25. Andò, B.; Baglio, S.; Graziani, S.; Marletta, V.; Dibilio, V.; Mostile, G.; Zappia, M. A Comparison among Different Strategies to Detect Potential Unstable Behaviors in Postural Sway. Sensors 2022, 22, 7106. [Google Scholar] [CrossRef]
  26. Mazumder, O.; Kundu, A.S.; Lenka, P.K.; Bhaumik, S. Ambulatory activity classification with dendogram-based support vector machine: Application in lower-limb active exoskeleton. Gait Posture 2016, 50, 53–59. [Google Scholar] [CrossRef]
  27. Andò, B.; Baglio, S.; Marletta, V.; Marrella, M.; Rajan, S.; Dibilio, V.; Mostile, G.; Zappia, M. A Neuro-Fuzzy-Based Sensing Approach for the Classification of Emulated Postural Instability. IEEE Sens. J. 2023, 23, 23866–23874. [Google Scholar] [CrossRef]
  28. Sabatini, A.M.; Mannini, A. Ambulatory Assessment of Instantaneous Velocity during Walking Using Inertial Sensor Measurements. Sensors 2016, 16, 2206. [Google Scholar] [CrossRef]
  29. Abujrida, H.; Agu, E.; Pahlavan, K. Machine learning-based motor assessment of Parkinson’s disease using postural sway, gait and lifestyle features on crowdsourced smartphone data. Biomed. Phys. Eng. Express 2020, 6, 035005. [Google Scholar] [CrossRef]
  30. Adın, V.; Zhang, Y.; Andò, B.; Oelmann, B.; Bader, S. Tiny Machine Learning for Real-Time Postural Stability Analysis. In Proceedings of the 2023 IEEE Sensors Applications Symposium (SAS), Ottawa, ON, Canada, 18–20 July 2023; pp. 1–6. [Google Scholar]
  31. Barkallah, E.; Freulard, J.; Otis, M.J.D.; Ngomo, S.; Ayena, J.C.; Desrosiers, C. Wearable Devices for Classification of Inadequate Posture at Work Using Neural Networks. Sensors 2017, 17, 2003. [Google Scholar] [CrossRef]
  32. Lim, H.; Kim, B.; Park, S. Prediction of Lower Limb Kinetics and Kinematics during Walking by a Single IMU on the Lower Back Using Machine Learning. Sensors 2019, 20, 130. [Google Scholar] [CrossRef] [PubMed]
  33. Greene, B.R.; McManus, K.; Ader, L.G.M.; Caulfield, B. Unsupervised Assessment of Balance and Falls Risk Using a Smartphone and Machine Learning. Sensors 2021, 21, 4770. [Google Scholar] [CrossRef]
  34. Prado, A.; Kwei, S.K.; Vanegas-Arroyave, N.; Agrawal, S.K. Continuous Identification of Freezing of Gait in Parkinson’s Patients Using Artificial Neural Networks and Instrumented Shoes. IEEE Trans. Med. Robot. Bionics 2021, 3, 554–562. [Google Scholar] [CrossRef]
  35. Kamran, F.; Harrold, K.; Zwier, J.; Carender, W.; Bao, T.; Sienko, K.H.; Wiens, J. Automatically evaluating balance using machine learning and data from a single inertial measurement unit. J. Neuroeng. Rehabil. 2021, 18, 114. [Google Scholar] [CrossRef]
  36. Bao, T.; Klatt, B.N.; Whitney, S.L.; Sienko, K.H.; Wiens, J. Automatically Evaluating Balance: A Machine Learning Approach. IEEE Trans. Neural Syst. Rehabil. Eng. 2019, 27, 179–186. [Google Scholar] [CrossRef]
  37. Palmerini, L.; Rocchi, L.; Mellone, S.; Valzania, F.; Chiari, L. Feature selection for accelerometer-based posture analysis in Parkinson’s disease. IEEE Trans. Inf. Technol. Biomed. 2011, 15, 481–490. [Google Scholar] [CrossRef] [PubMed]
  38. Sotirakis, C.; Su, Z.; Brzezicki, M.A.; Conway, N.; Tarassenko, L.; FitzGerald, J.J.; Antoniades, C.A. Identification of motor progression in Parkinson’s disease using wearable sensors and machine learning. NPJ Park. Dis. 2023, 9, 142. [Google Scholar] [CrossRef]
  39. Li, T.-H.S.; Kuo, P.-H.; Chen, L.-H.; Hung, C.-C.; Luan, P.-C.; Hsu, H.-P.; Chang, C.-H.; Hsieh, Y.-T.; Lin, W.-H. Fuzzy Double Deep Q-Network-Based Gait Pattern Controller for Humanoid Robots. IEEE Trans. Fuzzy Syst. 2022, 30, 147–161. [Google Scholar] [CrossRef]
  40. Chalvatzaki, G.; Koutras, P.; Hadfield, J.; Papageorgiou, X.S.; Tzafestas, C.S.; Maragos, P. LSTM-based Network for Human Gait Stability Prediction in an Intelligent Robotic Rollator. In Proceedings of the 2019 International Conference on Robotics and Automation (ICRA), Montreal, QC, Canada, 20–24 May 2019; pp. 4225–4232. [Google Scholar]
  41. Egan, S.; Brama, P.A.J.; Goulding, C.; McKeown, D.; Kearney, C.M.; McGrath, D. The Feasibility of Equine Field-Based Postural Sway Analysis Using a Single Inertial Sensor. Sensors 2021, 21, 1286. [Google Scholar] [CrossRef]
  42. Duong, T.T.; Uher, D.; Young, S.D.; Duong, T.; Sangco, M.; Cornett, K.; Montes, J.; Zanotto, D. Gaussian Process Regression for COP Trajectory Estimation in Healthy and Pathological Gait Using Instrumented Insoles. In Proceedings of the 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Prague, Czech Republic, 27 September–1 October 2021. [Google Scholar]
  43. Refai, M.I.M.; van Beijnum, B.-J.F.; Buurke, J.H.; Veltink, P.H. Portable Gait Lab: Instantaneous centre of mass velocity using three inertial measurement units. In Proceedings of the 2020 IEEE SENSORS, Rotterdam, The Netherlands, 25–28 October 2020; pp. 1–4. [Google Scholar]
  44. Pierleoni, P.; Belli, A.; Palma, L.; Paoletti, M.; Raggiunto, S.; Pinti, F. Postural stability evaluation using wearable wireless sensor. In Proceedings of the 2019 IEEE 23rd International Symposium on Consumer Technologies (ISCT), Ancona, Italy, 19–21 June 2019; pp. 256–260. [Google Scholar]
  45. Pinheiro, V.; Carmo, J.D.; Nascimento, F.d.O.; Miosso, C. System for the analysis of human balance based on accelerometers and support vector machines. Comput. Methods Programs Biomed. Updat. 2023, 4, 100123. [Google Scholar] [CrossRef]
  46. Yu, C.-H.; Yeh, C.-C.; Lu, Y.-F.; Lu, Y.-L.; Wang, T.-M.; Lin, F.Y.-S.; Lu, T.-W. Recurrent Neural Network Methods for Extracting Dynamic Balance Variables during Gait from a Single Inertial Measurement Unit. Sensors 2023, 23, 9040. [Google Scholar] [CrossRef] [PubMed]
  47. Duong, T.T.H.; Uher, D.; Young, S.D.; Farooquee, R.; Druffner, A.; Pasternak, A.; Kanner, C.; Fragala-Pinkham, M.; Montes, J.; Zanotto, D. Accurate COP Trajectory Estimation in Healthy and Pathological Gait Using Multimodal Instrumented Insoles and Deep Learning Models. IEEE Trans. Neural Syst. Rehabil. Eng. 2023, 31, 4801–4811. [Google Scholar] [CrossRef]
  48. Wantanajittikul, K.; Wiboonsuntharangkoon, C.; Chuatrakoon, B.; Kongsawasdi, S. Application of Machine Learning to Predict Trajectory of the Center of Pressure (COP) Path of Postural Sway Using a Triaxial Inertial Sensor. Sci. World J. 2022, 2022, 9483665. [Google Scholar] [CrossRef]
  49. He, B.; Shi, P.; Li, X.; Fan, M.; Deng, Z.; Yu, H. A pelvic support weight rehabilitation system tracing the human center of mass height. Sheng Wu Yi Xue Gong Cheng Xue Za Zhi 2022, 39, 175–184. [Google Scholar]
  50. Chebel, E.; Tunc, B. Deep neural network approach for estimating the three-dimensional human center of mass using joint angles. J. Biomech. 2021, 126, 110648. [Google Scholar] [CrossRef] [PubMed]
  51. Hnat, S.K.; Audu, M.L.; Triolo, R.J.; Quinn, R.D. Estimating Center of Mass Kinematics During Perturbed Human Standing Using Accelerometers. J. Appl. Biomech. 2021, 37, 415–424. [Google Scholar] [CrossRef] [PubMed]
  52. Lee, M.; Park, S. Estimation of Three-Dimensional Lower Limb Kinetics Data during Walking Using Machine Learning from a Single IMU Attached to the Sacrum. Sensors 2020, 20, 6277. [Google Scholar] [CrossRef]
  53. Wu, C.-C.; Chen, Y.-J.; Hsu, C.-S.; Wen, Y.-T.; Lee, Y.-J. Multiple Inertial Measurement Unit Combination and Location for Center of Pressure Prediction in Gait. Front. Bioeng. Biotechnol. 2020, 8, 566474. [Google Scholar] [CrossRef]
  54. Podobnik, J.; Kraljić, D.; Zadravec, M.; Munih, M. Centre of Pressure Estimation during Walking Using Only Inertial-Measurement Units and End-To-End Statistical Modelling. Sensors 2020, 20, 6136. [Google Scholar] [CrossRef]
  55. Choi, A.; Jung, H.; Mun, J.H. Single Inertial Sensor-Based Neural Networks to Estimate COM-COP Inclination Angle During Walking. Sensors 2019, 19, 2974. [Google Scholar] [CrossRef]
  56. Chen, V.C.; Chen, S.-W. Establishing the waist as the better location for attaching a single accelerometer to estimate center of pressure trajectories. Clin. Biomech. 2018, 60, 30–38. [Google Scholar] [CrossRef] [PubMed]
  57. Nataraj, R.; Audu, M.L.; Triolo, R.J. Center of mass acceleration feedback control of standing balance by functional neuromuscular stimulation against external perterbations. J. IEEE Trans. Biomed. Eng. 2013, 60, 10–19. [Google Scholar] [CrossRef]
  58. Burns, P.B.; Rohrich, R.J.; Chung, K.C.M. The Levels of Evidence and Their Role in Evidence-Based Medicine. Plast. Reconstr. Surg. 2011, 128, 305–310. [Google Scholar] [CrossRef]
  59. Kriegeskorte, N.; Golan, T. Neural network models and deep learning. Curr. Biol. 2019, 29, R231–R236. [Google Scholar] [CrossRef] [PubMed]
  60. Boukhennoufa, I.; Zhai, X.; Utti, V.; Jackson, J.; McDonald-Maier, K.D. Wearable sensors and machine learning in post-stroke rehabilitation assessment: A systematic review. Biomed. Signal Process. Control 2022, 71, 103197. [Google Scholar] [CrossRef]
  61. Slemenšek, J.; Fister, I.; Geršak, J.; Bratina, B.; van Midden, V.M.; Pirtošek, Z.; Šafarič, R. Human Gait Activity Recognition Machine Learning Methods. Sensors 2023, 23, 745. [Google Scholar] [CrossRef]
  62. Chen, M.; Wang, H.; Yu, L.; Yeung, E.H.K.; Luo, J.; Tsui, K.-L.; Zhao, Y. A Systematic Review of Wearable Sensor-Based Technologies for Fall Risk Assessment in Older Adults. Sensors 2022, 22, 6752. [Google Scholar] [CrossRef]
  63. Howcroft, J.; Kofman, J.; Lemaire, E.D. Review of fall risk assessment in geriatric populations using inertial sensors. J. Neuroeng. Rehabil. 2013, 10, 91. [Google Scholar] [CrossRef]
  64. Jiao, Y.; Hart, R.; Reading, S.; Zhang, Y. Systematic review of automatic post-stroke gait classification systems. Gait Posture 2024, 109, 259–270. [Google Scholar] [CrossRef]
  65. Bhongade, A.; Gupta, R.; Gandhi, T.K. Machine Learning-Based Gait Characterization Using Single IMU Sensor. In Proceedings of the 2022 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS), Greater Noida, India, 4–5 November 2022; pp. 263–266. [Google Scholar]
  66. Rashidi, H.H.; Albahra, S.; Robertson, S.; Tran, N.K.; Hu, B. Common statistical concepts in the supervised Machine Learning arena. Front. Oncol. 2023, 13, 1130229. [Google Scholar] [CrossRef]
  67. Liu, M.M.; Herzog, W.; Savelberg, H.H. Dynamic muscle force predictions from EMG: An artificial neural network approach. J. Electromyogr. Kinesiol. 1999, 9, 391–400. [Google Scholar] [CrossRef] [PubMed]
Sensors 24 07280 g001
Figure 1. Pyramid for artificial intelligence scientific evidence [24]. Adapted from “The artificial intelligence evidence-based medicine pyramid,” by Bellini, V., et al., 2023, World J Crit Care Med, 2023. 12(2): p. 89–91. Copyright 2023. Reprinted with permission.
Figure 1. Pyramid for artificial intelligence scientific evidence [24]. Adapted from “The artificial intelligence evidence-based medicine pyramid,” by Bellini, V., et al., 2023, World J Crit Care Med, 2023. 12(2): p. 89–91. Copyright 2023. Reprinted with permission.
Sensors 24 07280 g001
Sensors 24 07280 g002
Figure 2. PRISMA flow diagram.
Figure 2. PRISMA flow diagram.
Sensors 24 07280 g002
Sensors 24 07280 g003
Figure 3. Summary of the different ML models applied in the reviewed studies.
Figure 3. Summary of the different ML models applied in the reviewed studies.
Sensors 24 07280 g003
Sensors 24 07280 g004
Figure 4. Sensor placement locations and the percentage of studies included in this review that consider each position (percentages are based on counts of studies that included that specific sensor location and may not add up to 100%).
Figure 4. Sensor placement locations and the percentage of studies included in this review that consider each position (percentages are based on counts of studies that included that specific sensor location and may not add up to 100%).
Sensors 24 07280 g004
Sensors 24 07280 g005
Figure 5. Wearable sensor characteristics.
Figure 5. Wearable sensor characteristics.
Sensors 24 07280 g005
Table 1. Systematic review search query.
Table 1. Systematic review search query.
ParameterSearch Query
Artificial intelligence
algorithm		(“artificial intelligence” OR AI OR “machine learning” OR ML OR “long short-term memory” OR LSTM OR “artificial neural network” OR ANN OR “neural network” OR “support vector machine” OR SVM) AND
Wearable sensor(“wearable sensor” OR IMU OR “inertial measurement unit” OR accelerometer OR gyroscope OR magnetometer) AND
Biomechanical characteristics(“center of pressure” OR COP OR “center of mass” OR sway)
Table 2. Summary of study characteristics. UND: trial duration for collected biomechanical data was not reported. CoP: center of pressure; CoM: center of mass; LSTM: long short-term memory; IMU: inertial measurement unit.
Table 2. Summary of study characteristics. UND: trial duration for collected biomechanical data was not reported. CoP: center of pressure; CoM: center of mass; LSTM: long short-term memory; IMU: inertial measurement unit.
First Author (Year)Study Design Level of EvidenceDowns & Black Scale ScoreParticipant CharacteristicsTesting ConditionData Collection Trial DurationEstimated Mechanical Parameter ML Method UsedWearable Sensors Used and Locations
Yu, C. H. (2023) [46]Internal Validation studyIII2126 healthy individuals, 13 subjects with 25.46 +/− 2.37 years of age and 13 subjects with 72.75 +/− 6.68 years of ages Walking UNDCoM-CoP inclination angle (IA)4 Types of Recurrent Neural Network (RNN) models: unidirectional-LSTM, bi-LSTM, uni-GRU, and bidirectional -GRU1 IMU; located on sacrum
Duong, Ton T. H. (2023) [47]Internal Validation studyIII1538 (13 healthy individuals and 25 with genetically determined neuromuscular conditions), 10.1–29.8 yearsWalking 6 minCoPBidirectional-LSTM model and Gaussian Process Regression (GP)8 FSRs on the foot and 1 IMU located under the midfoot
Labrozzi, Gabrielle (2023) [13]Internal Validation studyIII155 healthy individuals, 24–37 yearsWalking and standing balanceUNDCoMBi-directional LSTM model12 IMUs; three located on the thoracic region, three on the mid-abdomen, one positioned anterolaterally on each thigh and shank, and one per foot
Wantanajittikul, Kittichai (2022) [48]Internal Validation studyIII1853 healthy individuals, an average of 46 yearsStanding balance30 sCoPLeast-square boosting (LSBoost), Bootstrap aggregation (Bagging), Support vector machine (SVM), Artificial neural network (ANN), and Gaussian process (GP)1 IMU; located on the lumbar region of the trunk
Tan, T. (2022) [21]Internal Validation studyIII1516 healthy individuals, 23.0 +/− 2.2 yearsRunning 100 stepsStrike Index (CoP)Convolutional Neural Network (CNN)1 IMU sensor; located on dorsum surface of shoe
He, B. (2022) [49]Internal Validation studyIII1411 subjects (10 healthy individuals, 1 subject 8 months post-stroke), 22.7 +/− 3.6 years and 64 years oldWalking60 sCoM Height (CoMH)Back propagation (BP) neural network and GRU network (a variant of recurrent neural network (RNN))3 IMUs; located on the calf, thigh, and pelvis
Chebel, E. (2021) [50]Internal Validation studyIII1722 healthy individuals, 22–26 yearsWalking and squat tasksUNDCoMDeep Neural Network (DNN)5 IMUs; located on lumbosacral joint, right/left shoulders, and right/left hips
Hnat, S. K. (2021) [51]Internal Validation studyIII145 healthy individuals, Median age of 23 yearsStanding balance with internal and external perturbationsUNDCoMArtificial Neural Network (ANN)10 triaxial accelerometers; located on the torso, sternum, navel, pelvis, anterior thighs, and anterior shanks
Lee, M. (2020) [52]Internal Validation studyIII1520 healthy individuals, 24.7 +/− 3.2 yearsWalking2 minCoPArtificial Neural Network (ANN)1 IMU; located on the Sacrum
Wu, C. C. (2020) [53]Internal Validation studyIII145 healthy individuals, 25 +/− 1.87 yearsWalking74 stepsCoPLSTM model4 IMUs; located on left toe, lateral and heel parts of shoe, and waist
Podobnik, J. (2020) [54]Internal Validation studyIII126 subjects (4 healthy individuals, 2 stroke patients), 32-55 yearsWalking30 sec standing; 3 and 7 min walking trialsCoPNon-linear Long-Short-Term Memory (LSTM) model7 IMUs; located on the pelvis, each thigh, shank, and foot
Choi, A. (2019) [55]Internal Validation studyIII1424 healthy individuals, 26.2 +/− 1.5 yearsWalkingUNDCoM-CoP IAFeed forward ANN model and LSTM memory model1 IMU; located on lumbar spine
Chen, Vincent (2018) [56]Internal Validation studyIII1510 healthy individuals, 24–31 yearsStanding balance40 sCoPNeural network (NN), genetic algorithm (GA), and adaptive network-based fuzzy inference system (ANFIS) 3 accelerometers; located on the upper trunk, waist, and lower thigh
Nataraj, Raviraj (2012) [57]Case studyIV111 individual with a thoracic-level spinal cord injury, Age UnknownStanding balanceUNDCoM accelerationLinear regression model2 (3-D) accelerometers; located on the pelvis and torso
Table 3. Summary of study results. AP: anterior–posterior, ML: medio–lateral, [H]: healthy controls, [Patho]: pathological subjects, GP: Gaussian process model; Bi-LSTM: bi-directional long short-term memory model; Bagging: bootstrap aggregation; LSBoost: least-square boosting; SVM: support vector machine; ANN: artificial neural network; GA: genetic algorism, ANFIS: adaptive network-based fuzzy inference system; DNN: deep neural network, Uni-LSTM: uni-directional long-short-term memory model; Uni-GRU: uni-directional gated recurrent unit; Bi--GRU: bi-directional gated recurrent unit; CNN: convolutional neural network; GRU: gated recurrent unit; BP neural network: Back propagation training neural network.
Table 3. Summary of study results. AP: anterior–posterior, ML: medio–lateral, [H]: healthy controls, [Patho]: pathological subjects, GP: Gaussian process model; Bi-LSTM: bi-directional long short-term memory model; Bagging: bootstrap aggregation; LSBoost: least-square boosting; SVM: support vector machine; ANN: artificial neural network; GA: genetic algorism, ANFIS: adaptive network-based fuzzy inference system; DNN: deep neural network, Uni-LSTM: uni-directional long-short-term memory model; Uni-GRU: uni-directional gated recurrent unit; Bi--GRU: bi-directional gated recurrent unit; CNN: convolutional neural network; GRU: gated recurrent unit; BP neural network: Back propagation training neural network.
First Author (Year)Testing ConditionEstimated Mechanical ParameterMachine Learning ModelRoot-Mean-square Error (RMSE) [cm or degrees]Normalized RMSE (NRMSE) [%]Spearman’s Rho Correlation CoefficientCorrelation CoefficientError Ratio (ER) [%]Jaccard Index [%]
Yu, C. H. (2023) [46]WalkingCoP-CoM Inclination Angle (IA)Uni-LSTMAP: ~0.75 degAP: ~4.5%----
ML: ~0.55 degML: ~6%----
Bi-LSTMAP: ~0.65 degAP: ~4%----
ML: ~0.50 degML: ~5.7%----
Uni-GRUAP: ~0.62 degAP: ~3.9%----
ML: ~0.49 degML: ~5.5%----
Bi-GRUAP: 0.61 (0.24) degAP: 3.82 (1.53) %----
ML: 0.46 (0.21) degML: 5.33 (3.76) %----
Duong, T. (2023) [47]WalkingCoPGPAP: 1.74 (0.39) cm [H]; 1.90 (0.48) [Patho]AP: 5.85(1.64) % [H]; 6.93 (2.32) % [Patho]-AP: 0.95 [H]; 0.86 [Patho]--
ML: 0.56 (0.09) cm [H]; 0.70 (0.12) [Patho]ML: 8.94(2.34) % [H]; 9.49 (1.63) % [Patho]-ML: 0.39 [H]; 0.33 [Patho]--
Bi-LSTMAP: 1.44 (0.29) cm [H]; 1.53 (0.39) cm [Patho]AP: 4.79(1.11) % [H]; 5.59 (1.83) % [Patho]-AP: 0.97 [H]; 0.91 [Patho]--
ML: 0.51 (0.10) cm [H]; 0.60 (0.11) cm [Patho]ML: 7.96(1.92) % [H]; 8.13 (1.41) % [Patho]-ML: 0.51 [H]; 0.47 [Patho]--
Labrozzi, Gabrielle (2023) [13]Walking and standingCoMBi-LSTMAP: 1.15 (0.80) cm [H]; 1.77 (0.62) [Patho]-----
ML: 1.44 (0.65) cm [H]; 2.91 (0.62) [Patho]-----
Wantanajittikul, K. (2022) [48]Standing variations (double stance, tandem, and single leg stance)CoPBagging--0.8894---
SVM0.8978
ANN0.8921
GP0.8934
LSBoostUND
Tan, T. (2022) [21]RunningStrike Index (CoP)CNN-6.9 (1.5) %-0.89 (0.09)--
He, B. (2022) [49]WalkingCoM Height (CoMH)GRU0.1408~0.1662 cm--0.906~0.922--
BP Neural Network0.1717~0.1835 cm--0.811~0.854--
Chebel, E. (2021) [50]Walking and squat tasksCoMDNNAP: 0.845 (1.28) cm-----
ML: 0.603 (0.24) cm
Hnat, S. K. (2021) [51]Standing with internal and external perturbationsCoMANNAP: 0.94 cmAP: 14.6%-0.83--
ML: 0.60 cmML: 22.0%
Lee, M. (2020) [52]WalkingCoPANN-AP: 8.22 (4.96) %-AP: 0.86 (0.30)--
ML: 19.54 (9.59)%ML: 0.01 (1.43)
Wu, C. C. (2020) [53]WalkingCoPLSTM-AP: 5.88 (0.96) %---93%
ML: 25.33 (2.35)%63%
Podobnik, J. (2020) [54]WalkingCoPNon-linear LSTMAP: 1.49 cm-----
ML: 0.90 cm
Choi, A. (2019) [55]WalkingCoP-CoM Inclination Angle (IA)LSTMAP: 1.97 (0.81) deg--0.92--
ML: 0.82 (0.16) deg0.96
ANNAP: 3.01 (0.18) deg0.81
ML: 1.27 (0.05) deg0.87
Chen, Vincent (2018) [56]Standing while following AP/ML excursion requestsCoPGA---AP: 0.95 (0.03)AP: 8.1 (2.5) %-
ML: 0.96 (0.03)ML: 7.3 (2.6) %
ANFISAP: 0.95 (0.02)AP: 8.2 (2.3) %
ML: 0.96 (0.03)ML: 7.3 (2.7) %
Neural NetworkAP: 0.95 (0.02)AP: 8.4 (2.3) %
ML: 0.96 (0.03)ML: 7.0 (2.6) %
Nataraj, R. (2013) [57]Standing with external perturbationsCoM AccelerationLinear Regression---AP: 0.972--
ML: 0.993
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Museck, I.J.; Brinton, D.L.; Dean, J.C. The Use of Wearable Sensors and Machine Learning Methods to Estimate Biomechanical Characteristics During Standing Posture or Locomotion: A Systematic Review. Sensors 2024, 24, 7280. https://doi.org/10.3390/s24227280

AMA Style

Museck IJ, Brinton DL, Dean JC. The Use of Wearable Sensors and Machine Learning Methods to Estimate Biomechanical Characteristics During Standing Posture or Locomotion: A Systematic Review. Sensors. 2024; 24(22):7280. https://doi.org/10.3390/s24227280

Chicago/Turabian Style

Museck, Isabelle J., Daniel L. Brinton, and Jesse C. Dean. 2024. "The Use of Wearable Sensors and Machine Learning Methods to Estimate Biomechanical Characteristics During Standing Posture or Locomotion: A Systematic Review" Sensors 24, no. 22: 7280. https://doi.org/10.3390/s24227280

APA Style

Museck, I. J., Brinton, D. L., & Dean, J. C. (2024). The Use of Wearable Sensors and Machine Learning Methods to Estimate Biomechanical Characteristics During Standing Posture or Locomotion: A Systematic Review. Sensors, 24(22), 7280. https://doi.org/10.3390/s24227280

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop