Machine-Learning-Based Prediction of Gait Events From EMG in Cerebral Palsy Children

Estimating gait parameters from sEMG signals using machine learning techniques under different power capacity of muscle

The gait analysis has been applied in many fields, such as the assessment of falling, force evaluation in sports, and gait disorder detection for neuromuscular diseases. Its main recording techniques include video cameras and wearable sensors. However, the present methods involve measuring surface electromyograms (sEMGs) to analyze muscle activities. The primary goal of this study is to estimate gait parameters under different power capacity of muscle by sEMGs measured from lower limbs. A self-made wireless device recorded sEMGs from two muscles of each foot, and GaitUp Physilog®5 sensors captured gait parameters from 18 participants under running as references. Four features including median frequency (MDF), waveform length (WL), standard deviation (SD), and sample entropy (SampEn), were extracted from the sEMG data. The analysis utilized three machine learning models (Random Forest, CatBoost, XGBoost), evaluated through various evaluation metrics. Additionally, 5-fold cross-validation was conducted to assess the influence of muscle fatigue on the estimation of gait parameters. The results show that all models successfully estimated 20 gait parameters, all showing a Pearson correlation coefficient (PCC) above 0.800. However, the performance of models significantly depends on the condition of muscle fatigue. This study represents a significant advancement in gait analysis, providing a comprehensive method for estimating gait parameters from sEMG signals, with important implications for mobile health applications.

Deep learning-based classification of hemiplegia and diplegia in cerebral palsy using postural control analysis

Cerebral palsy (CP) is a neurological condition that affects mobility and motor control, presenting significant challenges for accurate diagnosis, particularly in cases of hemiplegia and diplegia. This study proposes a method of classification utilizing Recurrent Neural Networks (RNNs) to analyze time series force data obtained via an AMTI platform. The proposed research focuses on optimizing these models through advanced techniques such as automatic parameter optimization and data augmentation, improving the accuracy and reliability in classifying these conditions. The results demonstrate the effectiveness of the proposed models in capturing complex temporal dynamics, with the Bidirectional Gated Recurrent Unit (BiGRU) and Long Short-Term Memory (LSTM) model achieving the highest performance, reaching an accuracy of 76.43%. These results outperform traditional approaches and offer a valuable tool for implementation in clinical settings. Moreover, significant differences in postural stability were observed among patients under different visual conditions, underscoring the importance of tailoring therapeutic interventions to each patient's specific needs.

Estimating Ground Reaction Forces from Gait Kinematics in Cerebral Palsy: A Convolutional Neural Network Approach

PURPOSE

While gait analysis is essential for assessing neuromotor disorders like cerebral palsy (CP), capturing accurate ground reaction force (GRF) measurements during natural walking presents challenges, particularly due to variations in gait patterns. Previous studies have explored GRF prediction using machine learning, but specific focus on patients with CP is lacking. This research aims to address this gap by predicting GRF using joint angles derived from marker data during gait in patients with CP, thereby suggesting a protocol for gait analysis without the need for force plates.

METHODS

The study employed an extensive dataset comprising both typically developed (TD) subjects (n = 132) and patients with CP (n = 622), captured using motion capture systems and force plates. Kinematic data included lower limb angles in three planes of motion, while GRF data encompassed three axes. A one-dimensional convolutional neural network model was designed to extract features from kinematic time series, followed by densely connected layers for GRF prediction. Evaluation metrics included normalized root mean squared error (nRMSE) and Pearson correlation coefficient (PCC).

RESULTS

GRFs of patients with CP were predicted with nRMSE values consistently below 20.13% and PCC scores surpassing 0.84. In the TD group, all GRFs were predicted with higher accuracy, showing nRMSE values lower than 12.65% and PCC scores exceeding 0.94.

CONCLUSION

The predictions considerably captured the patterns observed in the experimentally obtained GRFs. Despite limitations, including the absence of upper extremity kinematics data and the need for continuous model evolution, the study demonstrates the potential of machine learning in predicting GRFs in patients with CP, albeit with current prediction errors constraining immediate clinical applicability.

Lateral walking gait phase recognition for hip exoskeleton by denoising autoencoder-LSTM

Lateral resistance walk is an effective way to strengthen the abductor muscles of the hip. Accurate lateral walking gait recognition is the prerequisite for exoskeletons to be applied to lateral walking exercises. This paper proposes a denoising autoencoder-LSTM (DAE-LSTM) algorithm for lateral walking gait recognition. Nine sets of IMU data at three speeds and three strides of ten subjects were collected. Four lateral walking gait phases of narrow double support (NDS), guided foot swing (GFS), wide double support (WDS) and following leg swing (FLS) were recognized. The recognition performance of random forest (RF), support vector machine (SVM), k-nearest neighbors (KNN), neural networks (NN) and DAE-LSTM were compared. The average cross-subject recognition accuracy of DAE-LSTM was 90.2 %, which was higher than the other four models and previous work. For each frame of IMU data, the average recognition time of DAE-LSTM is 0.383 ms, which is 5.32 ms higher than the previous work. When the signal-to-noise ratio (SNR) is greater than 100:1, the accuracy of the DAE-LSTM model is higher than 90.0 %, and the accuracy of the other four models were less than 85 %. The results show that the proposed algorithm can achieve the requirements of recognition accuracy, model recognition time and model robustness for application in exoskeleton.

Artificial intelligence and machine learning approaches in cerebral palsy diagnosis, prognosis, and management: a comprehensive review

Cerebral palsy (CP) is a group of disorders that alters patients' muscle coordination, posture, and movement, resulting in a wide range of deformities. Cerebral palsy can be caused by various factors, both prenatal and postnatal, such as infections or injuries that damage different parts of the brain. As brain plasticity is more prevalent during childhood, early detection can help take the necessary course of management and treatments that would significantly benefit patients by improving their quality of life. Currently, cerebral palsy patients receive regular physiotherapies, occupational therapies, speech therapies, and medications to deal with secondary abnormalities arising due to CP. Advancements in artificial intelligence (AI) and machine learning (ML) over the years have demonstrated the potential to improve the diagnosis, prognosis, and management of CP. This review article synthesizes existing research on AI and ML techniques applied to CP. It provides a comprehensive overview of the role of AI-ML in cerebral palsy, focusing on its applications, benefits, challenges, and future prospects. Through an extensive examination of existing literature, we explore various AI-ML approaches, including but not limited to assessment, diagnosis, treatment planning, and outcome prediction for cerebral palsy. Additionally, we address the ethical considerations, technical limitations, and barriers to the widespread adoption of AI-ML for CP patient care. By synthesizing current knowledge and identifying gaps in research, this review aims to guide future endeavors in harnessing AI-ML for optimizing outcomes and transforming care delivery in cerebral palsy rehabilitation.

Application of Isokinetic Dynamometry Data in Predicting Gait Deviation Index Using Machine Learning in Stroke Patients: A Cross-Sectional Study

BACKGROUND

Three-dimensional gait analysis, supported by advanced sensor systems, is a crucial component in the rehabilitation assessment of post-stroke hemiplegic patients. However, the sensor data generated from such analyses are often complex and challenging to interpret in clinical practice, requiring significant time and complicated procedures. The Gait Deviation Index (GDI) serves as a simplified metric for quantifying the severity of pathological gait. Although isokinetic dynamometry, utilizing sophisticated sensors, is widely employed in muscle function assessment and rehabilitation, its application in gait analysis remains underexplored.

OBJECTIVE

This study aims to investigate the use of sensor-acquired isokinetic muscle strength data, combined with machine learning techniques, to predict the GDI in hemiplegic patients. This study utilizes data captured from sensors embedded in the Biodex dynamometry system and the Vicon 3D motion capture system, highlighting the integration of sensor technology in clinical gait analysis.

METHODS

This study was a cross-sectional, observational study that included a cohort of 150 post-stroke hemiplegic patients. The sensor data included measurements such as peak torque, peak torque/body weight, maximum work of repeated actions, coefficient of variation, average power, total work, acceleration time, deceleration time, range of motion, and average peak torque for both flexor and extensor muscles on the affected side at three angular velocities (60°/s, 90°/s, and 120°/s) using the Biodex System 4 Pro. The GDI was calculated using data from a Vicon 3D motion capture system. This study employed four machine learning models-Lasso Regression, Random Forest (RF), Support Vector regression (SVR), and BP Neural Network-to model and validate the sensor data. Model performance was evaluated using mean squared error (MSE), the coefficient of determination (R2), and mean absolute error (MAE). SHapley Additive exPlanations (SHAP) analysis was used to enhance model interpretability.

RESULTS

The RF model outperformed others in predicting GDI, with an MSE of 16.18, an R2 of 0.89, and an MAE of 2.99. In contrast, the Lasso Regression model yielded an MSE of 22.29, an R2 of 0.85, and an MAE of 3.71. The SVR model had an MSE of 31.58, an R2 of 0.82, and an MAE of 7.68, while the BP Neural Network model exhibited the poorest performance with an MSE of 50.38, an R2 of 0.79, and an MAE of 9.59. SHAP analysis identified the maximum work of repeated actions of the extensor muscles at 60°/s and 120°/s as the most critical sensor-derived features for predicting GDI, underscoring the importance of muscle strength metrics at varying speeds in rehabilitation assessments.

CONCLUSIONS

This study highlights the potential of integrating advanced sensor technology with machine learning techniques in the analysis of complex clinical data. The developed GDI prediction model, based on sensor-acquired isokinetic dynamometry data, offers a novel, streamlined, and effective tool for assessing rehabilitation progress in post-stroke hemiplegic patients, with promising implications for broader clinical application.

A systematic review on machine learning approaches in cerebral palsy research

Background

This review aims to explore advances in the field of cerebral palsy (CP) focusing on machine learning (ML) models. The objectives of this study is to analyze the advances in the application of ML models in the field of CP and to compare the performance of different ML algorithms in terms of their effectiveness in CP identification, classifying CP into its subtypes, prediction of abnormalities in CP, and its management. These objectives guide the review in examining how ML techniques are applied to CP and their potential impact on improving outcomes in CP research and treatment.

Methodology

A total of 20 studies were identified on ML for CP from 2013 to 2023. Search Engines used during the review included electronic databases like PubMed for accessing biomedical and life sciences, IEEE Xplore for technical literature in computer, Google Scholar for a broad range of academic publications, Scopus and Web of Science for multidisciplinary high impact journals. Inclusion criteria included articles containing keywords such as cerebral palsy, machine learning approaches, outcome response, identification, classification, diagnosis, and treatment prediction. Studies were included if they reported the application of ML techniques for CP patients. Peer reviewed articles from 2013 to 2023 were only included for the review. We selected full-text articles, clinical trials, randomized control trial, systematic reviews, narrative reviews, and meta-analyses published in English. Exclusion criteria for the review included studies not directly related to CP. Editorials, opinion pieces, and non-peer-reviewed articles were also excluded. To ensure the validity and reliability of the findings in this review, we thoroughly examined the study designs, focusing on the appropriateness of their methodologies and sample sizes. To synthesize and present the results, data were extracted and organized into tables for easy comparison. The results were presented through a combination of text, tables, and figures, with key findings emphasized in summary tables and relevant graphs.

Results

Random forest (RF) is mainly used for classifying movements and deformities due to CP. Support vector machine (SVM), decision tree (DT), RF, and K-nearest neighbors (KNN) show 100% accuracy in exercise evaluation. RF and DT show 94% accuracy in the classification of gait patterns, multilayer perceptron (MLP) shows 84% accuracy in the classification of CP children, Bayesian causal forests (BCF) have 74% accuracy in predicting the average treatment effect on various orthopedic and neurological conditions. Neural networks are 94.17% accurate in diagnosing CP using eye images. However, the studies varied significantly in their design, sample size, and quality of data, which limits the generalizability of the findings.

Conclusion

Clinical data are primarily used in ML models in the CP field, accounting for almost 47%. With the rise in popularity of machine learning techniques, there has been a rise in interest in developing automated and data-driven approaches to explore the use of ML in CP.

Accuracy of EMG linear envelope in identifying the peak of muscular activity during walking

BACKGROUND

The linear-envelope peak (LEP) of surface EMG signal is widely used in gait analysis to characterize muscular activity, especially in clinics.

RESEARCH QUESTION

This study is designed to evaluate LEP accuracy in identifying muscular activation and assessing activation timing during walking.

METHODS

Surface EMG signals from gastrocnemius lateralis (GL) and tibialis anterior (TA) were analyzed in 100 strides per subject (31 healthy subjects) during ground walking. Signals were full-wave rectified and low-pass filtered (cut-off frequency=5 Hz) to extract the linear envelope. LEP accuracy in identifying muscle activations and the associated error in peak detection were assessed by direct comparison with a reference method based on wavelet transform. LEP accuracy in identifying the timing of higher signalenergy levels was also assessed, increasing the reference-algorithm selectivity.

RESULTS

The detection error (percentage number of times when LEP falls outside the correspondent reference activation interval) is close to zero. Detection error increases up to 70% for intervals including only signal energy higher than 90% of energy peak. Mean absolute error (MAE, the absolute value of the distance between LEP timing and the correspondent actual timing of the sEMG-signal peak computed by reference algorithm) is 54.1±20.0 ms. Detection error and MAE are significantly higher (p<0.05) in TA data compared to GL signals. Differences among MAE values detected adopting different values for LE cut-off frequency are not statistically significant.

SIGNIFICANCE

LEP was found to be accurate in identifying the number of muscle activations during walking. However, the use of LEP to assess the timing of highest sEMG-signal energy (signal peak) should be considered carefully. Indeed, it could introduce a relevant inaccuracy in muscle-activation identification and peak-timing quantification. The type of muscle to analyze could also influence LEP performances, while the cut-off frequency chosen for envelope extraction appears to have a limited impact.

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

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

Adaptive neuro-fuzzy inference system model driven by the non-negative matrix factorization-extracted muscle synergy patterns to estimate lower limb joint movements

BACKGROUND AND OBJECTIVE

For patients with movement disorders, the main clinical focus is on exercise rehabilitation to help recover lost motor function, which is achieved by relevant assisted equipment. The basis for seamless control of the assisted equipment is to achieve accurate inference of the user's movement intentions in the human-machine interface. This study proposed a novel movement intention detection technology for estimating lower limb joint continuous kinematic variables following muscle synergy patterns, to develop applications for more efficient assisted rehabilitation training.

METHODS

This study recruited 16 healthy males and 16 male patients with symptomatic patellar tendinopathy (VISA-P: 59.1 ± 8.7). The surface electromyography of 12 muscles and lower limb joint kinematic and kinetic data from healthy subjects and patients during step-off landings from 30 cm-high stair steps were collected. We subsequently solved the preprocessed data based on the established recursive model of second-order differential equation to obtain the muscle activation matrix, and then imported it into the non-negative matrix factorization model to obtain the muscle synergy matrix. Finally, the lower limb neuromuscular synergy pattern was then imported into the developed adaptive neuro-fuzzy inference system non-linear regression model to estimate the human movement intention during this movement pattern.

RESULTS

Six muscle synergies were determined to construct the muscle synergy pattern driven ANFIS model. Three fuzzy rules were determined in most estimation cases. Combining the results of the four error indicators across the estimated variables indicates that the current model has excellent estimated performance in estimating lower limb joint movement. The estimation errors between the healthy (Angle: R2=0.98±0.03; Torque: R2=0.96±0.04) and patient (Angle: R2=0.98±0.02; Torque: R2=0.96±0.03) groups are consistent.

CONCLUSION

The proposed model of this study can accurately and reliably estimate lower limb joint movements, and the effectiveness will also be radiated to the patient group. This revealed that our models also have certain advantages in the recognition of motor intentions in patients with relevant movement disorders. Future work from this study can be focused on sports rehabilitation in the clinical field by achieving more flexible and precise movement control of the lower limb assisted equipment to help the rehabilitation of patients.

Multi-Channel FES Gait Rehabilitation Assistance System Based on Adaptive sEMG Modulation

Functional electrical stimulation (FES) can be used to stimulate the lower-limb muscles to provide walking assistance to stroke patients. However, the existing surface electromyography (sEMG)-based FES control methods mostly only consider a single muscle with a fixed stimulation intensity and frequency. This study proposes a multi-channel FES gait rehabilitation assistance system based on adaptive myoelectric modulation. The proposed system collects sEMG of the vastus lateralis muscle on the non-affected side to predict the sEMG values of four targeted lower-limb muscles on the affected side using a bidirectional long short-term memory (BILSTM) model. Next, the proposed system modulates the real-time FES output frequency for four targeted muscles based on the predicted sEMG values to provide muscle force compensation. Fifteen healthy subjects were recruited to participate in an offline model-building experiment conducted to evaluate the feasibility of the proposed BILSTM model in predicting the sEMG values. The experimental results showed that the [Formula: see text] value of the best-obtained prediction result reached 0.85 using the BILSTM model, which was significantly higher than that using traditional prediction methods. Moreover, two patients after stroke were recruited in the online assisted-walking experiment to verify the effectiveness of the proposed walking-assistance system. The experimental results showed that the activation of the target muscles of the patients was higher after FES, and the gait movement data were significantly different before and after FES. The proposed system can be effectively applied to walking assistance for stroke patients, and the experimental results can provide new ideas and methods for sEMG-controlled FES rehabilitation applications.

Machine learning-based prediction of joint moments based on kinematics in patients with cerebral palsy

Joint moments during gait provide valuable information for clinical decision-making in patients with cerebral palsy (CP). Joint moments are calculated based on ground reaction forces (GRF) using inverse dynamics models. Obtaining GRF from patients with CP is challenging. Typically developed (TD) individuals' joint moments were predicted from joint angles using machine learning, but no such study has been conducted on patients with CP. Accordingly, we aimed to predict the dorsi-plantar flexion, knee flexion-extension, hip flexion-extension, and hip adduction-abduction moments based on the trunk, pelvis, hip, knee, and ankle kinematics during gait in patients with CP and TD individuals using one-dimensional convolutional neural networks (CNN). The anonymized retrospective gait data of 329 TD (26 years ± 14, mass: 70 kg ± 15, height: 167 cm ± 89) and 917 CP (17 years ± 9, mass:47 kg ± 19, height:153 cm ± 36) individuals were evaluated and after applying inclusion-exclusion criteria, 132 TD and 622 CP patients with spastic diplegia were selected. We trained specific CNN models and evaluated their performance using isolated test subject groups based on normalized root mean square error (nRMSE) and Pearson correlation coefficient (PCC). Joint moments were predicted with nRMSE between 18.02% and 13.58% for the CP and between 12.55% and 8.58% for the TD groups, whereas with PCC between 0.85 and 0.93 for the CP and between 0.94 and 0.98 for the TD groups. Machine learning-based joint moment prediction from kinematics could replace conventional moment calculation in CP patients in the future, but the current level of prediction errors restricts its use for clinical decision-making today.

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

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

Flexible Machine Learning Algorithms for Clinical Gait Assessment Tools

The current gold standard of gait diagnostics is dependent on large, expensive motion-capture laboratories and highly trained clinical and technical staff. Wearable sensor systems combined with machine learning may help to improve the accessibility of objective gait assessments in a broad clinical context. However, current algorithms lack flexibility and require large training datasets with tedious manual labelling of data. The current study tests the validity of a novel machine learning algorithm for automated gait partitioning of laboratory-based and sensor-based gait data. The developed artificial intelligence tool was used in patients with a central neurological lesion and severe gait impairments. To build the novel algorithm, 2% and 3% of the entire dataset (567 and 368 steps in total, respectively) were required for assessments with laboratory equipment and inertial measurement units. The mean errors of machine learning-based gait partitions were 0.021 s for the laboratory-based datasets and 0.034 s for the sensor-based datasets. Combining reinforcement learning with a deep neural network allows significant reduction in the size of the training datasets to <5%. The low number of required training data provides end-users with a high degree of flexibility. Non-experts can easily adjust the developed algorithm and modify the training library depending on the measurement system and clinical population.

Automatic Hemiplegia Type Detection (Right or Left) Using the Levenberg-Marquardt Backpropagation Method

Hemiplegia affects a significant portion of the human population. It is a condition that causes motor impairment and severely reduces the patient’s quality of life. This paper presents an automatic system for identifying the hemiplegia type (right or left part of the body is affected). The proposed system utilizes the data taken from patients and healthy subjects using the accelerometer sensor from the RehaGait mobile gait analysis system. The collected data undergo a pre-processing procedure followed by a feature extraction stage. The extracted features are then sent to a neural network trained by the Levenberg-Marquardt backpropagation (LM-BP) algorithm. The experimental part of this research involved creating a custom-created dataset containing entries taken from ten healthy and twenty non-healthy subjects. The data were taken from seven different sensors placed in specific areas of the subjects’ bodies. These sensors can capture a three-dimensional (3D) signal using the accelerometer, magnetometer, and gyroscope device types. The proposed system used the signals taken from the accelerometers, which were split into 2-sec windows. The proposed system achieved a classification accuracy of 95.12% and was compared with fourteen commonly used machine learning approaches.