Liangliang Xiang
Alan Wang
Yaodong Gu
Liang Zhao1*
Vickie Shim
Justin Fernandez- 1Faculty of Sports Science, Ningbo University, Ningbo, China
- 2Research Academy of Grand Health, Ningbo University, Ningbo, China
- 3Auckland Bioengineering Institute, The University of Auckland, Auckland, New Zealand
- 4Faculty of Medical and Health Sciences, The University of Auckland, Auckland, New Zealand
- 5Department of Engineering Science, Faculty of Engineering, The University of Auckland, Auckland, New Zealand
With the emergence of wearable technology and machine learning approaches, gait monitoring in real-time is attracting interest from the sports biomechanics community. This study presents a systematic review of machine learning approaches in running biomechanics using wearable sensors. Electronic databases were retrieved in PubMed, Web of Science, SPORTDiscus, Scopus, IEEE Xplore, and ScienceDirect. A total of 4,068 articles were identified via electronic databases. Twenty-four articles that met the eligibility criteria after article screening were included in this systematic review. The range of quality scores of the included studies is from 0.78 to 1.00, with 40% of articles recruiting participant numbers between 20 and 50. The number of inertial measurement unit (IMU) placed on the lower limbs varied from 1 to 5, mainly in the pelvis, thigh, distal tibia, and foot. Deep learning algorithms occupied 57% of total machine learning approaches. Convolutional neural networks (CNN) were the most frequently used deep learning algorithm. However, the validation process for machine learning models was lacking in some studies and should be given more attention in future research. The deep learning model combining multiple CNN and recurrent neural networks (RNN) was observed to extract different running features from the wearable sensors and presents a growing trend in running biomechanics.
Introduction
Machine learning approaches have been widely utilized in gait biomechanics studies in the past decades (Ferber et al., 2016; Halilaj et al., 2018; Xiang et al., 2022a). Most runners frequently suffer running-related injuries in the lower limb and foot (van Gent et al., 2007; Murr and Pierce, 2019). Based on learning-driven sensor data, machine learning and deep learning could provide gait monitoring in real-time, recommendations for running shoes (Koska and Maiwald, 2020; Young et al., 2020), and suggestions for running injury protection (Girka et al., 2020; Matijevich et al., 2020; Dempster et al., 2021). It is crucial to know how machine learning techniques are implemented in lower limb running biomechanics by exploring wearable sensor data.
Traditional gait biomechanical analysis methods use statistical hypothesis tests based on discrete variables and summary metrics, such as the mean peak angles (Taylor et al., 2013; Halilaj et al., 2018; Dixon et al., 2019). A priori selection of features and sufficient background knowledge are essential to conducting further analysis and that may decrease the objectivity (Phinyomark et al., 2015, 2018). Consequently, these statistical tools cannot explain the complexity of multi-variables and process data-intense tasks (Halilaj et al., 2018). In the era of big data, machine learning as a growing data science method can process and analyze the large amounts of gait biomechanics data, for instance, wearable sensor data, and achieve state-of-the-art performance (Ferber et al., 2016).
Machine learning incorporates supervised learning and unsupervised learning. In the realm of biomechanics, it has been utilized for dimensionality reduction of high-dimensional data (Phinyomark et al., 2015; Watari et al., 2018a), gait classification (Fukuchi et al., 2011; Taylor et al., 2013; Clermont et al., 2017), pathology detection (Bennetts et al., 2013; Christian et al., 2016; Li et al., 2020), and human activity recognition (Ordóñez and Roggen, 2016; Ignatov, 2018; Ihianle et al., 2020). Principle component analysis (PCA) as an unsupervised learning method is commonly used for feature extraction before training (Wu and Wang, 2008; Taylor et al., 2013; Phinyomark et al., 2014; Clermont et al., 2017; Mei et al., 2020; Suda et al., 2020). Clermont et al. (2017) classified competitive and recreational runners based on lower limb kinematics data using the support vector machine (SVM). Machine learning algorithms can also discriminate runners' experience level and gender via assessing running's spatiotemporal parameters (Clermont et al., 2019b). Bennetts et al. (2013) identified the typical peak plantar pressure distributions utilizing k-means clustering.
Lower limb gait biomechanical studies are typically limited to the laboratory-based setting (Xiang et al., 2022b), such as measuring knee joint angles and moments using an optical motion capture system and force plates (Liu et al., 2020). Inertial measurement unit (IMU) is portable, lightweight, and low-cost, and can be used in an unconstrained environment (Fong and Chan, 2010; Ahamed et al., 2019). Wearable sensors have gained popularity for gait analysis in recent years. A recent systematic review from Camomilla et al. (2018) summarized the growing trend of utilizing wearable inertial sensors in the field of biomechanical analysis and activity classification. Picerno (2017) compared different approaches to evaluating lower limb joint kinematics by using IMU sensors. Deep learning is a subset of machine learning algorithms based on artificial neural networks (ANN). In recent years, it has been widely used in computer vision (Lee et al., 2009; Sermanet et al., 2013), speech recognition (Sainath et al., 2015), and medical image analysis (Ker et al., 2017; Shen et al., 2017). Deep neural network structures have advantages in processing time-series sensor data and require lower computational cost than traditional machine learning approaches. Integrated with deep learning algorithms, wearable sensors can be utilized for pattern recognition (Cust et al., 2019) and biomechanical variable prediction (Stetter et al., 2019; Hernandez et al., 2021) without experiment and environmental limitations.
Hu et al. (2018) found that long short-term memory (LSTM) recurrent neural networks (RNN) can detect surface- and age-related differences in walking gait based on a single wearable IMU sensor. Ronao and Cho (2016) showed that the accelerometer and gyroscope sensor data adopting deep convolutional neural networks (CNN) achieved high accuracy for human activity recognition. Ordóñez and Roggen (2016) presented a deep learning framework of the convolutional and LSTM (DeepConvLSTM) for accurate human activity recognition. Convolutional layers act as feature extractors capturing spatial domain features from one-dimensional sensor data, while recurrent layers are used to extract temporal domain features (Hernandez et al., 2021). The performance improves by fusing the accelerometer, gyroscope, and magnetic field sensors compared to the acceleration or acceleration and angular velocity data (Ordóñez and Roggen, 2016).
IMU sensors are heavily used technologies in distance running (Zrenner et al., 2020). Using inertial sensor-based data during running, one can classify different conditions and predict kinetical variables (Clermont et al., 2019b; Pogson et al., 2020). Jogging gait phase and period were detected and identified by lower extremity placed accelerometers and gyroscopes using machine learning algorithms (Mannini and Sabatini, 2012; Zdravevski et al., 2017). Kobsar et al. (2014) utilized PCA to classify training background from running experience. In a study by Pogson et al. (2020), PCA and multilayer perceptron (MLP) were adopted to predict ground reaction force (GRF) from trunk acceleration. Running speed conditions and running environments were assessed and classified using SVM with an IMU sensor placed on the participants' lower back during data collection (Benson et al., 2018b, 2020).
In 2018, Benson et al. (2018a) reviewed the progress of IMU sensors in gait analysis. O'Reilly et al. (2018) evaluated the lower limb exercise detection accuracy of the wearable inertial sensor. However, that review retrieved articles till 2017 and did not include running gait. A systematic review from Farrahi et al. (2019) revealed that machine learning techniques could predict activity type and intensity based on raw acceleration data. Fong and Chan (2010) estimated the use of wearable sensors in lower limb biomechanics. The approaches to assessing lower limb joint kinematics by using wearable sensors have been summarized by Picerno (2017). A systematic review in 2019 illustrated wearable inertial sensors' performance in sport-specific movement recognition using machine learning and deep learning approaches (Rapp et al., 2021). Although there is a growing trend concerning machine learning in lower limb running biomechanics, particularly deep learning algorithms in wearable inertial sensor studies, there are few studies on the accuracy of machine learning approaches as utilized in lower limb running biomechanics integrating wearable inertial sensors. Furthermore, no compelling evidence illustrated the application scenes of different machine learning algorithms, the requirements of sensor placement based on the research goal, and how the model was assessed and validated in lower limb running biomechanics.
Based on the currently existing knowledge gap, the purpose of this initial study is to conduct a systematic review regarding machine learning and deep learning approaches used in running biomechanics and was limited to the wearable sensors placed in lower limbs. By investigating the eligible studies, we hope: (1) to elaborate on different kinds of machine learning techniques used in running biomechanics and its performance; (2) to recommend suitable sensor placement locations to obtain decent accelerations or angular velocities or other biomechanical variables; (3) improve predictive accuracy for the related studies in gait analysis in the future.
Methods
Search Strategy
This systematic review followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) recommendations (Moher et al., 2009). The protocol for this systematic review was registered on INPLASY (NO. 202210083). Electronic databases were retrieved in PubMed, Web of Science, SPORTDiscus, Scopus, IEEE Xplore, and ScienceDirect by one reviewer (L.X.) to identify original research articles published up to May 2021 (from 2000). The paper screening was conducted by two investigators independently (L.X. and A.W.). After an initial search, relative article information was input to Rayyan QCRI (Ouzzani et al., 2016) for duplicate removal, study screening, and identification. A backward search was conducted on the studies included. One study was added from the reference lists of included studies to review the flow for further paper screening. Non-English articles and conference proceedings, and dissertations were excluded. The retrieve strategy and limit conditions are shown in Table 1. Four different categories were used to identify relevant studies: wearable inertial sensor, machine learning and deep learning, lower limb, and running. By using Boolean operation, the retrieved studies at least contain one keyword in the full field. The flow diagram of the paper search and screen process is presented in Figure 1.
Table 1. Electronic databases retrieve strategy.
Figure 1. PRISMA flow diagram for original research articles' searching and screening process.
Eligibility Criteria
Articles were selected based on the PICO principle (participants, intervention, comparisons, and outcomes). Our research identified studies that utilized machine learning and deep learning approaches and wearable inertial sensors to assess running biomechanics. For the included studies, wearable sensors must be placed in the participants' lower limbs. Only inertial sensors were selected in this review, including accelerometer, gyroscope, and magnetometer. Studies were removed if the statistical regression model rather than the machine learning approach was employed for the prediction. Energy expenditure and sprint performance assessment studies were not considered in this study. Studies using machine learning and wearable sensors only evaluating walking gait and human activity recognition were excluded.
Quality Assessment
The methodological quality of the included studies was assessed by the modified QualSyst Assessment Tool for quantitative studies (Kmet et al., 2004). A total of ten items were identified in this scale for aspects of the research question, study design, statistical analysis, sample size, and results reporting. Each item was scored as “yes” = 2, “partial” = 1, or “no” = 0 given the degree of the specific criteria that were met. The risk of bias was initially evaluated by three investigators (L.X., A.W., and J.F.), independently and confirmed by two reviewers (A.W. and J.F.). Each study's score was calculated by summing 10 items' scores and dividing by the total score (Kandula et al., 2016; Capiau et al., 2020). The evaluation questions include If the question or objective is clearly described? Is design evident and appropriate to answer the study question? If the method of subject selection or source of information/input variables is described and appropriate? Is the subject's characteristics, input variables, or information sufficiently described? Is the outcome well defined and robust to measurement/misclassification bias? If means of assessment are reported? Is the sample size appropriate? If analysis is described and appropriate? If some estimate of variance is reported for the main results? Are the results reported in sufficient detail? Do the results support the conclusions?
Data Extraction
Information was extracted from the included 24 studies based on participants and sensor characteristics, and machine learning approaches. Participant information included participant numbers, gender, age, type, and running speed during the data collection. Sensor characteristics contained sensor number, placement in the low limb, brand, and sampling frequency for data collection. For machine learning and deep learning, specific approaches, percentage of training and test dataset, cross-validation (CV) methods, data preprocessing, predictors, response, ground true reference, evaluation, and performance were extracted from the included literature.
Results
Search Results
A total of 4,068 articles were identified via the electronic databases retrieve, and one additional study was found from the reference lists. Then, 405 duplicate articles were removed; 119 studies remained for the full-text evaluation after the screening of the study's title and abstract; 95 articles were excluded based on the exclusion criteria. Finally, 24 articles were included in this systematic review.
Quality Assessment
As shown in Table 2, the range of quality scores of the included studies is from 0.78 to 1.00, and the mean score is 0.91. The main sources of bias are the small sample size and unclear subject characteristics. Seven studies vaguely reported the study design. Statistical analysis was not included in eight studies, but machine learning approaches' performance and accuracy were reported.
Table 2. Methodological quality assessment by the modified QualSyst quality appraisal tool.
Participants' Characteristics
The majority of participants were healthy subjects, except for three studies that contained patellofemoral pain participants or subjects with running-related injuries (Table 3) (Watari et al., 2018a,b; Rapp et al., 2021). The healthy subjects included both experienced runners and novice runners. The sample size was from 6 to 580, and three articles with a sample size below ten (Figure 2A) (Ahamed et al., 2018; Ngoh et al., 2018; Wouda et al., 2018). Approximately 40% of articles recruited participants with numbers between 20 and 50. In six studies, only male subjects were considered (Ngoh et al., 2018; Wouda et al., 2018; Stetter et al., 2019, 2020; Gholami et al., 2020; Hernandez et al., 2021). Age characteristics in three articles were not reported (Young et al., 2020; Johnson et al., 2021; Rapp et al., 2021). Several studies tested various speeds, except the running speed not mentioned in two studies (Dixon et al., 2019; Tan et al., 2019).
Table 3. Participants and wearable inertial sensor specifications.
Figure 2. Characteristic information: (A) sample size; (B) the number of sensors; (C) types of machine learning algorithms; (D) machine learning approaches; (E) purpose of machine learning.
Sensor Information
The number of IMUs sensors placed on the lower limbs was 1, 2, 3, and 5 (as shown in Figure 2B). Pelvis, thigh, shank, distal tibia, and foot were frequently selected as the placement location. Sensor(s) was placed on or in the running shoes in four studies (Ngoh et al., 2018; Zrenner et al., 2018; Gholami et al., 2020; Koska and Maiwald, 2020). Three types of IMUs sensors are utilized, including commercial sensors, custom-built sensors, and virtual sensors. All studies contain acceleration data with the range of accelerometers from ±6 to ±50 g. Gyroscope data were not contained in nine articles (Watari et al., 2018a,b; Dixon et al., 2019; Komaris et al., 2019; Tan et al., 2019; Derie et al., 2020; Matijevich et al., 2020; Johnson et al., 2021; Robberechts et al., 2021) and for most of the studies, magnetometer data were not incorporated in the IMU sensor (Watari et al., 2018a,b; Zrenner et al., 2018; Dixon et al., 2019; Komaris et al., 2019; Stetter et al., 2019, 2020; Tan et al., 2019; Derie et al., 2020; Koska and Maiwald, 2020; Matijevich et al., 2020; Young et al., 2020; Hernandez et al., 2021; Johnson et al., 2021; Rapp et al., 2021; Robberechts et al., 2021).
Machine Learning and Deep Learning
Approaches and Purposes
As depicted in Figures 2C–E, Deep learning algorithms occupied 57% of total machine learning approaches, compared with 43% of traditional machine learning algorithms. CNN was the most frequently used algorithm (seven times), followed by ANN and long short-term memory (LSTM). The DeepConvLSTM model was recently applied to lower limb joint prediction from linear acceleration and angular velocity data of the IMUs sensor (Hernandez et al., 2021). Gradient boosting (GB) algorithm included gradient boosting regression tree (Derie et al., 2020) and decision tree (Dixon et al., 2019; Liu et al., 2020; Young et al., 2020). There are two unsupervised learning algorithms [i.e., hierarchical cluster analysis (Watari et al., 2018b) and K-means clustering (Koska and Maiwald, 2020)] adopted for the clustering tasks. Eight machine learning approaches were used to classify different conditions. Regression tasks were the primary intention in 15 articles.
Lower limb running kinematics and kinetics were predicted using wearable inertial sensors combined with regression algorithms, including joint angles, forces, and moments and GRF (as shown in Table 4). Different types of outdoor terrain (Dixon et al., 2019), inclinations of the running surface (Ahamed et al., 2019), and environmental weather conditions (Ahamed et al., 2018) were detected and classified in three studies. The accuracy of gait event and spatiotemporal parameter detections was also tested (Zrenner et al., 2018; Tan et al., 2019; Liu et al., 2020; Robberechts et al., 2021). Two studies aimed at the running pattern or level classification (Clermont et al., 2019a; Liu et al., 2020). One study divided patellofemoral pain patients into two subgroups based on running gait (Watari et al., 2018b). The accuracy of identifying the effectiveness of exercise treatment for patellofemoral pain patients was evaluated with an SVM classifier by the same team (Watari et al., 2018a). PCA was adopted for dimensionality reduction in these two articles. According to foot posture and foot strike pattern, the ensemble deep learning model could accurately assess and recommend running shoe types to runners with an average accuracy of 94.6% (Young et al., 2020). Comfortable and uncomfortable running shoes also could be identified from the heel's sagittal plane angular velocity data using the k-nearest neighbors' classifier (KNN) (Koska and Maiwald, 2020).
Table 4. The detailed machine learning approaches.
Predictors of Machine Learning Model
For predictors in the studies, time-series data from the trial-axis accelerometer were chosen as attributes in 15 articles (Watari et al., 2018a,b; Zrenner et al., 2018; Dixon et al., 2019; Komaris et al., 2019; Stetter et al., 2019, 2020; Tan et al., 2019, 2020; Gholami et al., 2020; Liu et al., 2020; Young et al., 2020; Hernandez et al., 2021; Johnson et al., 2021; Rapp et al., 2021). Most of the studies used deep learning neural network algorithms for prediction. Continuing trial-axis angular velocity data were considered in eight studies (Zrenner et al., 2018; Stetter et al., 2019, 2020; Liu et al., 2020; Tan et al., 2020; Young et al., 2020; Hernandez et al., 2021; Rapp et al., 2021). Sagittal plane angular velocity (Koska and Maiwald, 2020) and anterior-posterior (Ngoh et al., 2018), and vertical (Wouda et al., 2018) acceleration were investigated in three studies. Discrete biomechanical variables measured from wearable sensors were also selected as input (Watari et al., 2018a,b; Clermont et al., 2019a). Three studies extracted the statistical features in raw sensor signals as the attributes (Derie et al., 2020; Matijevich et al., 2020; Robberechts et al., 2021). The predicted outcomes of the regression algorithms were compared with the ground truth reference data, which was measured from the optical motion capture system and force plate. For the collection of continuing time-series data, the force-instrumented treadmill was employed.
Cross-Validation and Evaluation
Several studies had no validation process for the machine learning algorithms in the methodology section (Ahamed et al., 2018, 2019; Watari et al., 2018a; Zrenner et al., 2018; Dixon et al., 2019; Derie et al., 2020; Koska and Maiwald, 2020; Matijevich et al., 2020; Stetter et al., 2020; Tan et al., 2020; Young et al., 2020; Johnson et al., 2021). In one study, KNN was trained in different proportions of the dataset, but trained models were not validated or tested (Koska and Maiwald, 2020). Half of the studies used leave-one-subject-out cross-validation (LOSOCV) methods (Wouda et al., 2018; Zrenner et al., 2018; Ahamed et al., 2019; Komaris et al., 2019; Stetter et al., 2019, 2020; Derie et al., 2020; Gholami et al., 2020; Liu et al., 2020; Matijevich et al., 2020; Tan et al., 2020; Robberechts et al., 2021). For the regression assignments, Pearson's correlation coefficient (R2), root-mean-squared error (RMSE), mean error (ME), and mean absolute error (MAE) were utilized to assess the model's performance. Accuracy, F1 score, precision, recall, confusion matrix, Matthews correlation coefficient (MCC), and receiver operating characteristic curves (ROC) were employed to evaluate the classification problems.
Discussion
This systematic study evaluated the use of wearable inertial sensors combined with machine learning and deep learning algorithms in the field of low limb running biomechanics. The pelvis, tibia, and foot were common locations for the sensor placement, and two sensors were most frequently adopted. Simulated IMU signals were also explored by converting marker trajectories into accelerations via numerical differentiation (Johnson et al., 2021; Rapp et al., 2021). It was found that the use of IMU sensors with machine learning approaches emerged recently (from 2018). The performance of assessing joint angles, forces, moments and GRF, and identifying and classifying multiple conditions were investigated. Furthermore, processing time-series data from IMUs using deep learning algorithms to predict lower limb biomechanics and classification tasks are becoming increasingly prevalent.
Machine Learning-Based Methods for Evaluating Running Biomechanics
Musculoskeletal models and kinematic chain models are physics-based approaches introduced to calculate gait kinematics and kinetics from IMU sensors (Karatsidis et al., 2017; Picerno, 2017; Dorschky et al., 2019). Subject-specific anthropometric data, however, are mandatory to scale the musculoskeletal model (Stetter et al., 2020). This process could inevitably cause inaccuracy (Faber et al., 2016; Ancillao et al., 2018). Kinematic chain modeling needs to capture the kinematic behaviors of main body segments by attaching one sensor to each segment (Wouda et al., 2018). Therefore, it takes a longer time for the experimental setup, requires multiple sensors, and constrains gait movement (Ngoh et al., 2018). In contrast to physics-based models, data-driven approaches used fewer sensors and built-up and optimized model parameters by training the model using part of the data rather than requiring any prior knowledge of the model (Wouda et al., 2018; Stetter et al., 2020).
The accuracy of predicted knee joint angles using CNN is higher than hip and ankle angles, even though the sensor data were obtained from the foot (Gholami et al., 2020). This means the sensor location may not be the most critical factor for the estimation of lower joint kinematics. Another factor, for instance, the flexibility of joints, may also affect the accuracy. Tenforde et al. (2020) found that tibial acceleration is associated with the GRF matrix in injured runners. However, Matijevich et al. (2019, 2020) clarified that the vertical average loading rate (VALR) during running is not strongly correlated with peak tibial force. Miller et al. (2019) using tibial acceleration combined with ANN predicted vertical GRF across multiple speeds [RMSE = 0.16 body weight (BW), R2 = 0.97]. Linear regression with least absolute shrinkage and selection operator regularization (LASSO) regression algorithm also showed promising results for evaluating peak tibial bone load from pressure and foot-mounted IMU data (Matijevich et al., 2020). These findings highlight that machine learning and deep learning algorithms could successfully predict response even though the relationship between inputs and output is still unclear and these issues can be explained from the perspective of data-driven approaches.
Model Assessment
Only two public datasets of running biomechanics (Komaris et al., 2019; Tan et al., 2019) were employed for analysis within the included studies, and the sample size is limited (below thirty). Generally, the CNN model sets 2–3 hidden layers, the LSTM usually contains two layers, and the ANN layer varies from 1 to 3 layers. The number of neurons in each hidden layer ranged from 10 to 250. Adam was the most frequently used optimizer, but those should be determined by hyperparameter tuning for the best combination (Hernandez et al., 2021). The predicted error of sagittal joint angles from the simulated IMU data has been shown to decrease (Dorschky et al., 2020). This might be due to the simulated data being smoother than the measured data (with less noise effect) (Rapp et al., 2021). Different validation or test approaches were compared to check their influence on the predicted accuracy (Wouda et al., 2018; Derie et al., 2020; Gholami et al., 2020; Liu et al., 2020). This review noticed that some studies did not validate the model's performance, which should be improved in future research. Best accuracy was achieved by splitting one subject's trials into both training and test procedures. The data not seen during testing were vital to assess machine learning models' performance and improve confidence in its practice (Halilaj et al., 2018). This intra-participant method could reduce the reliability and practicality in real-world applications (Ahamed et al., 2019; Derie et al., 2020; Gholami et al., 2020). Nested k-fold CV could conduct both hyperparameter tuning and evaluation based on inner and outer loops (Hernandez et al., 2021). It is recommended to use the LOSOCV or Nested k-fold CV method to validate or test the model's performance. The leave-one-trial-out cross-validation (LOTOCV) or random train test split approach will be only suitable for validating or testing the subject-specific machine learning model (Ahamed et al., 2019; Derie et al., 2020).
The Practice of Deep Learning
The deep learning technique takes time-series data into the input and has high computational efficiency compared with traditional machine learning approaches. Both CNN and RNN are popular tools utilized in lower extremity running biomechanics. A new approach (Ordóñez and Roggen, 2016) called DeepConvLSTM developed from human activity recognition classification was adopted to predict lower limb kinematics and shows state-of-the-art accuracy (Hernandez et al., 2021). Spatiotemporal features of multiple wearable sensors were extracted through CNN and RNN layers in these neural networks. Ihianle et al. (2020) found that DeepConvLSTM achieved the highest accuracy for the classification of jogging and running from multiple daily activities, compared to CNN and LSTM algorithms based on wearable sensor data. However, currently, only lower limb kinematics has been estimated using the DeepConvLSTM model. This review demonstrated that running kinetics is the most investigated prediction task. Based on the knowledge gap, it could be valuable to explore this algorithm's ability to predict lower limb kinetics during running. However, the current applications of the deep learning technique suffer from one main pitfall. Data-driven deep learning algorithms require large datasets for model training, but such datasets are scarce in running biomechanics compared with gait datasets in walking, which incorporate ground truth values from the motion capture system and wearable sensor data. On the other hand, data augmentation or transfer learning techniques can be considered to improve the model's generalization and performance on the limited training dataset (Komaris et al., 2019; Rapp et al., 2021).
Recommendations for Future Studies
According to the information from these included articles and analysis, there are several limitations in the reviewed previous studies. The following directions are identified and should be reviewed for future research regarding machine learning in running biomechanics by using wearable inertial sensor data:
• The sample size to garner sensor data for assessing running biomechanics is recommended to be larger than 20.
• One or two sensors are enough to obtain predictors inputting to the machine learning and deep learning models.
• The tibia, foot, and pelvis are frequently employed locations for sensors' attaching.
• Subject-independent (inter-participant) methods should be used to test the performance of the machine learning model (Derie et al., 2020; Liu et al., 2020; Tan et al., 2020; Hernandez et al., 2021).
• Raw acceleration data is recommended as input to capture the variabilities of spatial and temporal features (Dixon et al., 2019).
• Acceleration data are among the most commonly adopted wearable inertial data in lower extremity running biomechanics.
• Deep learning approaches could be more suitable for dealing with time-series data from wearable sensor data.
• Hyperparameter tuning is not only beneficial for picking the best combination of the model's parameters but also for the model's structure selection (Hernandez et al., 2021).
• Dividing training, validation, and testing datasets rigorously and presenting data not seen before during testing are essential to assess machine learning models' generalization and improve confidence in their practice (Halilaj et al., 2018).
• CNN, ANN, LSTM, DeepConvLSTM, MLP, and GB are popular algorithms to process IMUs data.
• The implementation of DeepConvLSTM for exploring wearable sensor signals in the field of lower limb running biomechanics is generally promising (Hernandez et al., 2021).
• Data augmentation or transfer learning approaches provide us with a novel viewpoint on running biomechanics, given the scarcity of data currently accessible in the field (Komaris et al., 2019; Rapp et al., 2021).
Conclusion
This study reviewed the current practice and trend in the realm of lower extremity biomechanics during running. Machine learning approaches, especially deep learning approaches, have rapidly arisen in recent years due to wearable technology improvements in gait analysis. Machine learning algorithms showed state-of-the-art predictability for processing wearable inertial data. However, in the future, the validation procedure for machine learning models should receive increased emphasis. A deep learning model combining resemble CNN and RNN should be utilized to extract different running features from the IMUs sensor. Investigating both upper limb and lower limb biomechanics for future studies would be interesting and worthwhile as running is a whole-body action.
Data Availability Statement
The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding authors.
Author Contributions
LX conceived the study idea, conducted study screening, and drafted the manuscript. AW, YG, LZ, VS, and JF conceived the study idea, assisted in revising the manuscript, and reviewed the first and final versions of the manuscript. All authors contributed to the article and agreed to the submitted version of the manuscript.
Funding
This work was supported by the National Key R&D Program of China (2018YFF0300905), the Key R&D Program of Zhejiang Province, China under Grant 2021C03130; Public Welfare Science and Technology Project of Ningbo, China under Grant 2021S133; Zhejiang Province Science Fund for Distinguished Young Scholars under Grant R22A021199; K. C. Wong Magna Fund in Ningbo University. LX would like to thank the support from China Scholarship Council (CSC).
Conflict of Interest
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher's Note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
Ahamed, N. U., Kobsar, D., Benson, L., Clermont, C., Kohrs, R., Osis, S. T., et al. (2018). Using wearable sensors to classify subject-specific running biomechanical gait patterns based on changes in environmental weather conditions. PLoS ONE 13, e0203839. doi: 10.1371/journal.pone.0203839
Ahamed, N. U., Kobsar, D., Benson, L. C., Clermont, C. A., Osis, S. T., and Ferber, R. (2019). Subject-specific and group-based running pattern classification using a single wearable sensor. J. Biomech. 84, 227–233. doi: 10.1016/j.jbiomech.2019.01.001
Ancillao, A., Tedesco, S., Barton, J., and O'Flynn, B. (2018). Indirect measurement of ground reaction forces and moments by means of wearable inertial sensors: a systematic review. Sensors 18, 2564. doi: 10.3390/s18082564
Bennetts, C. J., Owings, T. M., Erdemir, A., Botek, G., and Cavanagh, P. R. (2013). Clustering and classification of regional peak plantar pressures of diabetic feet. J. Biomech. 46, 19–25. doi: 10.1016/j.jbiomech.2012.09.007
Benson, L. C., Clermont, C. A., Bošnjak, E., and Ferber, R. (2018a). The use of wearable devices for walking and running gait analysis outside of the lab: a systematic review. Gait Posture 63, 124–138. doi: 10.1016/j.gaitpost.2018.04.047
Benson, L. C., Clermont, C. A., and Ferber, R. (2020). New considerations for collecting biomechanical data using wearable sensors: the effect of different running environments. Front. Bioeng. Biotechnol. 8, 86. doi: 10.3389/fbioe.2020.00086
Benson, L. C., Clermont, C. A., Osis, S. T., Kobsar, D., and Ferber, R. (2018b). Classifying running speed conditions using a single wearable sensor: optimal segmentation and feature extraction methods. J. Biomech. 71, 94–99.s. doi: 10.1016/j.jbiomech.2018.01.034
Camomilla, V., Bergamini, E., Fantozzi, S., and Vannozzi, G. (2018). Trends supporting the in-field use of wearable inertial sensors for sport performance evaluation: a systematic review. Sensors 18, 873. doi: 10.3390/s18030873
Capiau, A., Foubert, K., Van der Linden, L., Walgraeve, K., Hias, J., Spinewine, A., et al. (2020). Medication counselling in older patients prior to hospital discharge: a systematic review. Drugs Aging 37, 635–655. doi: 10.1007/s40266-020-00780-z
Christian, J., Kröll, J., Strutzenberger, G., Alexander, N., Ofner, M., and Schwameder, H. (2016). Computer aided analysis of gait patterns in patients with acute anterior cruciate ligament injury. Clin. Biomech. 33, 55–60. doi: 10.1016/j.clinbiomech.2016.02.008
Clermont, C. A., Benson, L. C., Edwards, W. B., Hettinga, B. A., and Ferber, R. (2019a). New considerations for wearable technology data: changes in running biomechanics during a marathon. J. Appl. Biomech. 35, 401–409. doi: 10.1123/jab.2018-0453
Clermont, C. A., Benson, L. C., Osis, S. T., Kobsar, D., and Ferber, R. (2019b). Running patterns for male and female competitive and recreational runners based on accelerometer data. J. Sports Sci. 37, 204–211. doi: 10.1080/02640414.2018.1488518
Clermont, C. A., Osis, S. T., Phinyomark, A., and Ferber, R. (2017). Kinematic gait patterns in competitive and recreational runners. J. Appl. Biomech. 33, 268–276. doi: 10.1123/jab.2016-0218
Cust, E. E., Sweeting, A. J., Ball, K., and Robertson, S. (2019). Machine and deep learning for sport-specific movement recognition: a systematic review of model development and performance. J. Sports Sci. 37, 568–600. doi: 10.1080/02640414.2018.1521769
Dempster, J., Dutheil, F., and Ugbolue, U. C. (2021). The prevalence of lower extremity injuries in running and associated risk factors: a systematic review. Phys. Activity Health 5, 133–145. doi: 10.5334/paah.109
Derie, R., Robberechts, P., Van den Berghe, P., Gerlo, J., De Clercq, D., Segers, V., et al. (2020). Tibial acceleration-based prediction of maximal vertical loading rate during overground running: a machine learning approach. Front. Bioeng. Biotechnol. 8, 33. doi: 10.3389/fbioe.2020.00033
Dixon, P. C., Schütte, K. H., Vanwanseele, B., Jacobs, J. V., Dennerlein, J. T., Schiffman, J. M., et al. (2019). Machine learning algorithms can classify outdoor terrain types during running using accelerometry data. Gait Posture 74, 176–181. doi: 10.1016/j.gaitpost.2019.09.005
Dorschky, E., Nitschke, M., Martindale, C. F., van den Bogert, A. J., Koelewijn, A. D., and Eskofier, B. M. (2020). CNN-based estimation of sagittal plane walking and running biomechanics from measured and simulated inertial sensor data. Front. Bioeng. Biotechnol. 8, 604. doi: 10.3389/fbioe.2020.00604
Dorschky, E., Nitschke, M., Seifer, A.-K., van den Bogert, A. J., and Eskofier, B. M. (2019). Estimation of gait kinematics and kinetics from inertial sensor data using optimal control of musculoskeletal models. J. Biomech. 95, 109278. doi: 10.1016/j.jbiomech.2019.07.022
Faber, G. S., Chang, C. C., Kingma, I., Dennerlein, J. T., and Van Dieën, J. H. (2016). Estimating 3D L5/S1 moments and ground reaction forces during trunk bending using a full-body ambulatory inertial motion capture system. J. Biomech. 49, 904–912. doi: 10.1016/j.jbiomech.2015.11.042
Farrahi, V., Niemelä, M., Kangas, M., Korpelainen, R., and Jämsä, T. (2019). Calibration and validation of accelerometer-based activity monitors: a systematic review of machine-learning approaches. Gait Posture 68, 285–299. doi: 10.1016/j.gaitpost.2018.12.003
Ferber, R., Osis, S. T., Hicks, J. L., and Delp, S. L. (2016). Gait biomechanics in the era of data science. J. Biomech. 49, 3759–3761. doi: 10.1016/j.jbiomech.2016.10.033
Fong, D. T. P., and Chan, Y. Y. (2010). The use of wearable inertial motion sensors in human lower limb biomechanics studies: a systematic review. Sensors 10, 11556–11565. doi: 10.3390/s101211556
Fukuchi, R. K., Eskofier, B. M., Duarte, M., and Ferber, R. (2011). Support vector machines for detecting age-related changes in running kinematics. J. Biomech. 44, 540–542. doi: 10.1016/j.jbiomech.2010.09.031
Gholami, M., Napier, C., and Menon, C. (2020). Estimating lower extremity running gait kinematics with a single accelerometer: a deep learning approach. Sensors 20, 2939. doi: 10.3390/s20102939
Girka, A., Kulmala, J. P., and Äyrämö, S. (2020). Deep learning approach for prediction of impact peak appearance at ground reaction force signal of running activity. Comput. Methods Biomech. Biomed. Eng. 23, 1052–1059. doi: 10.1080/10255842.2020.1786072
Halilaj, E., Rajagopal, A., Fiterau, M., Hicks, J. L., Hastie, T. J., and Delp, S. L. (2018). Machine learning in human movement biomechanics: best practices, common pitfalls, and new opportunities. J. Biomech. 81, 1–11. doi: 10.1016/j.jbiomech.2018.09.009
Hernandez, V., Dadkhah, D., Babakeshizadeh, V., and Kuli,ć, D. (2021). Lower body kinematics estimation from wearable sensors for walking and running: a deep learning approach. Gait Posture 83, 185–193. doi: 10.1016/j.gaitpost.2020.10.026
Hu, B., Dixon, P. C., Jacobs, J. V., Dennerlein, J. T., and Schiffman, J. M. (2018). Machine learning algorithms based on signals from a single wearable inertial sensor can detect surface- and age-related differences in walking. J. Biomech. 71, 37–42. doi: 10.1016/j.jbiomech.2018.01.005
Ignatov, A (2018). Real-time human activity recognition from accelerometer data using Convolutional Neural Networks. Appl. Soft Comput. J. 62, 915–922. doi: 10.1016/j.asoc.2017.09.027
Ihianle, I. K., Nwajana, A. O., Ebenuwa, S. H., Otuka, R. I., Owa, K., and Orisatoki, M. O. (2020). A deep learning approach for human activities recognition from multimodal sensing devices. IEEE Access 8, 179028–179038. doi: 10.1109/ACCESS.2020.3027979
Johnson, W. R., Mian, A., Robinson, M. A., Verheul, J., Lloyd, D. G., and Alderson, J. A. (2021). Multidimensional ground reaction forces and moments from wearable sensor accelerations via deep learning. IEEE Trans. Biomed. Eng. 68, 289–297. doi: 10.1109/TBME.2020.3006158
Kandula, T., Park, S. B., Cohn, R. J., Krishnan, A. V., and Farrar, M. A. (2016). Pediatric chemotherapy induced peripheral neuropathy: a systematic review of current knowledge. Cancer Treat. Rev. 50, 118–128. doi: 10.1016/j.ctrv.2016.09.005
Karatsidis, A., Bellusci, G., Schepers, H. M., De Zee, M., Andersen, M. S., and Veltink, P. H. (2017). Estimation of ground reaction forces and moments during gait using only inertial motion capture. Sensors 17, 75. doi: 10.3390/s17010075
Ker, J., Wang, L., Rao, J., and Lim, T. (2017). Deep learning applications in medical image analysis. IEEE Access 6, 9375–9389. doi: 10.1109/ACCESS.2017.2788044
Kmet, L. M., Cook, L. S., and Lee, R. C. (2004). Standard Quality Assessment Criteria for Evaluating Primary Research Papers from a Variety of Fields. doi: 10.7939/R37M04F16
Kobsar, D., Osis, S. T., Hettinga, B. A., and Ferber, R. (2014). Classification accuracy of a single tri-axial accelerometer for training background and experience level in runners. J. Biomech. 47, 2508–2511. doi: 10.1016/j.jbiomech.2014.04.017
Komaris, D. S., Pérez-Valero, E., Jordan, L., Barton, J., Hennessy, L., O'flynn, B., et al. (2019). Predicting three-dimensional ground reaction forces in running by using artificial neural networks and lower body kinematics. IEEE Access 7, 156779–156786. doi: 10.1109/ACCESS.2019.2949699
Koska, D., and Maiwald, C. (2020). A time-series based framework for exploring the unknown effects of shoe comfort-induced biomechanical adaptations. Footwear Sci. 12, 113–122. doi: 10.1080/19424280.2020.1734866
Lee, H., Grosse, R., Ranganath, R., and Ng, A. Y. (2009). “Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations,” in Proceedings of the 26th Annual International Conference on Machine Learning (Montreal, QC), 609–616.
Li, Y., Mache, M. A., and Todd, T. A. (2020). Automated identification of postural control for children with autism spectrum disorder using a machine learning approach. J. Biomech. 113, 110073. doi: 10.1016/j.jbiomech.2020.110073
Liu, Q., Mo, S., Cheung, V. C. K., Cheung, B. M. F., Wang, S., Chan, P. P. K., et al. (2020). Classification of runners' performance levels with concurrent prediction of biomechanical parameters using data from inertial measurement units. J. Biomech. 112, 110072. doi: 10.1016/j.jbiomech.2020.110072
Mannini, A., and Sabatini, A. M. (2012). Gait phase detection and discrimination between walking-jogging activities using hidden Markov models applied to foot motion data from a gyroscope. Gait Posture 36, 657–661. doi: 10.1016/j.gaitpost.2012.06.017
Matijevich, E. S., Branscombe, L. M., Scott, L. R., and Zelik, K. E. (2019). Ground reaction force metrics are not strongly correlated with tibial bone load when running across speeds and slopes: implications for science, sport and wearable tech. PLoS ONE 14, e0210000. doi: 10.1371/journal.pone.0210000
Matijevich, E. S., Scott, L. R., Volgyesi, P., Derry, K. H., and Zelik, K. E. (2020). Combining wearable sensor signals, machine learning and biomechanics to estimate tibial bone force and damage during running. Hum. Mov. Sci. 74, 102690. doi: 10.1016/j.humov.2020.102690
Mei, Q., Gu, Y., Xiang, L., Yu, P., Gao, Z., Shim, V., et al. (2020). Foot shape and plantar pressure relationships in shod and barefoot populations. Biomech. Model. Mechanobiol. 19, 1211–1224. doi: 10.1007/s10237-019-01255-w
Miller, A. C., Aljohani, M., Kim, H., and Kipp, K. (2019). “Neural network method to predicting stance-phase ground reaction force in distance runners,” in Proceedings of the 37th International Society of Biomechanics in Sport Conference (Oxford), 21–25.
Moher, D., Liberati, A., Tetzlaff, J., Altman, D. G., and Group, T. P. (2009). Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. PLoS Med. 6, e1000097. doi: 10.1371/journal.pmed.1000097
Murr, S., and Pierce, B. (2019). How aging impacts runners' goals of lifelong running. Phys. Act. Heal. 3, 71–81. doi: 10.5334/paah.42
Ngoh, K. J.-H., Gouwanda, D., Gopalai, A. A., and Zheng, C. Y. (2018). Estimation of vertical ground reaction force during running using neural network model and uniaxial accelerometer. J. Biomech. 76, 269–273. doi: 10.1016/j.jbiomech.2018.06.006
Ordóñez, F. J., and Roggen, D. (2016). Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity recognition. Sensors 16, 115. doi: 10.3390/s16010115
O'Reilly, M., Caulfield, B., Ward, T., Johnston, W., and Doherty, C. (2018). Wearable inertial sensor systems for lower limb exercise detection and evaluation: a systematic review. Sport. Med. 48, 1221–1246. doi: 10.1007/s40279-018-0878-4
Ouzzani, M., Hammady, H., Fedorowicz, Z., and Elmagarmid, A. (2016). Rayyan-a web and mobile app for systematic reviews. Syst. Rev. 5, 210. doi: 10.1186/s13643-016-0384-4
Phinyomark, A., Hettinga, B. A., Osis, S., and Ferber, R. (2015). Do intermediate- and higher-order principal components contain useful information to detect subtle changes in lower extremity biomechanics during running? Hum. Mov. Sci. 44, 91–101. doi: 10.1016/j.humov.2015.08.018
Phinyomark, A., Hettinga, B. A., Osis, S. T., and Ferber, R. (2014). Gender and age-related differences in bilateral lower extremity mechanics during treadmill running. PLoS ONE 9, e0105246. doi: 10.1371/journal.pone.0105246
Phinyomark, A., Petri, G., Ibáñez-Marcelo, E., Osis, S. T., and Ferber, R. (2018). Analysis of big data in gait biomechanics: current trends and future directions. J. Med. Biol. Eng. 38, 244–260. doi: 10.1007/s40846-017-0297-2
Picerno, P (2017). 25 years of lower limb joint kinematics by using inertial and magnetic sensors: a review of methodological approaches. Gait Posture 51, 239–246. doi: 10.1016/j.gaitpost.2016.11.008
Pogson, M., Verheul, J., Robinson, M. A., Vanrenterghem, J., and Lisboa, P. (2020). A neural network method to predict task- and step-specific ground reaction force magnitudes from trunk accelerations during running activities. Med. Eng. Phys. 78, 82–89. doi: 10.1016/j.medengphy.2020.02.002
Rapp, E., Shin, S., Thomsen, W., Ferber, R., and Halilaj, E. (2021). Estimation of kinematics from inertial measurement units using a combined deep learning and optimization framework. J. Biomech. 116, 110229. doi: 10.1016/j.jbiomech.2021.110229
Robberechts, P., Derie, R., Van den Berghe, P., Gerlo, J., De Clercq, D., Segers, V., et al. (2021). Predicting gait events from tibial acceleration in rearfoot running: a structured machine learning approach. Gait Posture 84, 87–92. doi: 10.1016/j.gaitpost.2020.10.035
Ronao, C. A., and Cho, S. B. (2016). Human activity recognition with smartphone sensors using deep learning neural networks. Expert Syst. Appl. 59, 235–244. doi: 10.1016/j.eswa.2016.04.032
Sainath, T. N., Kingsbury, B., Saon, G., Soltau, H., Mohamed, A., Dahl, G., et al. (2015). Deep convolutional neural networks for large-scale speech tasks. Neural networks 64, 39–48. doi: 10.1016/j.neunet.2014.08.005
Sermanet, P., Eigen, D., Zhang, X., Mathieu, M., Fergus, R., and LeCun, Y. (2013). Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv Prepr. arXiv1312.6229. doi: 10.48550/arXiv.1312.6229
Shen, D., Wu, G., and Suk, H.-I. (2017). Deep learning in medical image analysis. Annu. Rev. Biomed. Eng. 19, 221–248. doi: 10.1146/annurev-bioeng-071516-044442
Stetter, B. J., Krafft, F. C., Ringhof, S., Stein, T., and Sell, S. (2020). A machine learning and wearable sensor based approach to estimate external knee flexion and adduction moments during various locomotion tasks. Front. Bioeng. Biotechnol. 8, 9. doi: 10.3389/fbioe.2020.00009
Stetter, B. J., Ringhof, S., Krafft, F. C., Sell, S., and Stein, T. (2019). Estimation of knee joint forces in sport movements using wearable sensors and machine learning. Sensors 19, 1–12. doi: 10.3390/s19173690
Suda, E. Y., Watari, R., Matias, A. B., and Sacco, I. C. N. (2020). Recognition of foot-ankle movement patterns in long-distance runners with different experience levels using support vector machines. Front. Bioeng. Biotechnol. 8, 576. doi: 10.3389/fbioe.2020.00576
Tan, H., Aung, N. N., Tian, J., Chua, M. C. H., and Yang, Y. O. (2019). Time series classification using a modified LSTM approach from accelerometer-based data: a comparative study for gait cycle detection. Gait Posture 74, 128–134. doi: 10.1016/j.gaitpost.2019.09.007
Tan, T., Strout, Z. A., and Shull, P. B. (2020). Accurate impact loading rate estimation during running via a subject-independent convolutional neural network model and optimal IMU placement. IEEE J. Biomed. Heal. Informatics 25, 1215–1222. doi: 10.1109/JBHI.2020.3014963
Taylor, P., Eskofier, B. M., and Federolf, P. (2013). Marker-based classification of young-elderly gait pattern differences via direct PCA feature extraction and SVMs. Comput. Methods Biomech. Biomed. Engin. 16, 435–442. doi: 10.1080/10255842.2011.624515
Tenforde, A. S., Hayano, T., Jamison, S. T., Outerleys, J., and Davis, I. S. (2020). Tibial acceleration measured from wearable sensors is associated with loading rates in injured runners. PMR 12, 679–684. doi: 10.1002/pmrj.12275
van Gent, R. N., Siem, D., Van Middelkoop, M., Van Os, A. G., Bierma-Zeinstra, S. M. A., and Koes, B. W. (2007). Incidence and determinants of lower extremity running injuries in long distance runners: a systematic review. Br. J. Sports Med. 41, 469–480. doi: 10.1136/bjsm.2006.033548
Watari, R., Osis, S., and Ferber, R. (2018a). Use of baseline pelvic acceleration during running for classifying response to muscle strengthening treatment in patellofemoral pain: A preliminary study. Clin. Biomech. 57, 74–80. doi: 10.1016/j.clinbiomech.2018.06.010
Watari, R., Osis, S. T., Phinyomark, A., and Ferber, R. (2018b). Runners with patellofemoral pain demonstrate sub-groups of pelvic acceleration profiles using hierarchical cluster analysis: AN exploratory cross-sectional study. BMC Musculoskelet. Disord. 19, 1–11. doi: 10.1186/s12891-018-2045-3
Wouda, F. J., Giuberti, M., Bellusci, G., Maartens, E., Reenalda, J., van Beijnum, B. J. F., et al. (2018). Estimation of vertical ground reaction forces and sagittal knee kinematics during running using three inertial sensors. Front. Physiol. 9, 218. doi: 10.3389/fphys.2018.00218
Wu, J., and Wang, J. (2008). PCA-based SVM for automatic recognition of gait patterns. J. Appl. Biomech. 24, 83–87. doi: 10.1123/jab.24.1.83
Xiang, L., Gu, Y., Mei, Q., Wang, A., Shim, V., and Fernandez, J. (2022a). Automatic classification of barefoot and shod populations based on the foot metrics and plantar pressure patterns. Front. Bioeng. Biotechnol. 10, 843204. doi: 10.3389/fbioe.2022.843204
Xiang, L., Mei, Q., Wang, A., Shim, V., Fernandez, J., and Gu, Y. (2022b). Evaluating function in the hallux valgus foot following a 12-week minimalist footwear intervention: a pilot computational analysis. J. Biomech. 132, 110941. doi: 10.1016/j.jbiomech.2022.110941
Young, F., Coulby, G., Watson, I., Downs, C., Stuart, S., and Godfrey, A. (2020). Just find it: The Mymo approach to recommend running shoes. IEEE Access 8, 109791–109800. doi: 10.1109/ACCESS.2020.3002075
Zdravevski, E., Stojkoska, B. R., Standl, M., and Schulz, H. (2017). Automatic machine-learning based identification of jogging periods from accelerometer measurements of adolescents under field conditions. PLoS ONE 12, e0184216. doi: 10.1371/journal.pone.0184216
Zrenner, M., Gradl, S., Jensen, U., Ullrich, M., and Eskofier, B. M. (2018). Comparison of different algorithms for calculating velocity and stride length in running using inertial measurement units. Sensors 18, 4194. doi: 10.3390/s18124194
Keywords: gait, wearable sensor, machine learning, deep learning, running, lower limb
Citation: Xiang L, Wang A, Gu Y, Zhao L, Shim V and Fernandez J (2022) Recent Machine Learning Progress in Lower Limb Running Biomechanics With Wearable Technology: A Systematic Review. Front. Neurorobot. 16:913052. doi: 10.3389/fnbot.2022.913052
Received: 05 April 2022; Accepted: 04 May 2022;
Published: 02 June 2022.
Edited by:
Orazio Gambino, University of Palermo, ItalyReviewed by:
Zeki Akyildiz, Gazi University, TurkeyTyler A. Wood, Northern Illinois University, United States
Ravi Shekhar Tiwari, Chadura Tech, India
Marcus Chun Jin Tan, National University Health System, Singapore
Copyright © 2022 Xiang, Wang, Gu, Zhao, Shim and Fernandez. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Yaodong Gu, Z3V5YW9kb25nJiN4MDAwNDA7bmJ1LmVkdS5jbg==; Liang Zhao, emhhb2xpYW5nJiN4MDAwNDA7bmJ1LmVkdS5jbg==