Personalized EMG preprocessing and normalization for musculoskeletal simulation

Abstract

Background

Accurate interpretation of electromyography (EMG) signals is essential for reliable control of musculoskeletal (MS) models in biomechanics and rehabilitation applications. Conventional preprocessing methods may not account for subject-specific signal characteristics and task-related muscle function.

Objective

This study aimed to develop and validate an adaptive and personalized EMG preprocessing pipeline to enhance the physiological accuracy of EMG-driven musculoskeletal models during elbow flexion-extension tasks.

Methods

EMG signals from six upper limb muscles were recorded using a Delsys system while participants performed elbow flexion-extension movements. The signals were preprocessed using individualized spectral filtering and a dual-stage normalization approach. First, dynamic maximum voluntary contraction (MVC) based min–max normalization was applied to standardize signal amplitudes. Second, functional weighting was used to scale each muscle's activation based on its biomechanical contribution to the movement. The processed signals were used as input to an OpenSim elbow model, and resulting joint kinematics were compared to reference motion data captured by an Xsens system.

Results

The EMG-driven OpenSim model showed strong agreement with the Xsens data, with correlation coefficients exceeding 0.98 and root mean square error (RMSE) values below 8°. While a minor systematic offset was observed, joint angle trajectories remained consistent and physiologically plausible across trials.

Conclusion

The proposed subject-specific EMG preprocessing pipeline enhances the accuracy and interpretability of biomechanical models. Future research should explore adaptive signal alignment techniques and AI-based processing methods to improve model robustness in dynamic and wearable scenarios.

Keywords

musculoskeletal simulation EMG signal normalization OpenSim muscle weighting adaptive filtering

1 Introduction

Electromyography (EMG)-driven musculoskeletal models use measured muscle electrical activity to predict muscle forces and joint moments. These models have shown promise for understanding neuromuscular control and improving assistive device interfaces.¹ In rehabilitation technology, an EMG-driven model of the upper limb can enable user-intent detection and personalized assistance, for example by controlling exoskeletons or prostheses based on the patient's own muscle signals, performing rehabilitation exercises or choosing the most appropriate program.² In rehabilitation technology, EMG-driven upper limb models facilitate user-intent detection and personalized assistance. Recent studies demonstrate their effectiveness in controlling exoskeletons and prostheses. For instance, Samarakoon et al. (2025) reported a wearable LSTM–based exoskeleton that accurately predicted multi–joint user intent (96.2 % accuracy), while Lee et al. (2024) developed a soft exoskeleton capable of real–time intention detection (∼500 ms latency, ∼96 % accuracy).³ Cisnal et al. (2023) further showed that EMG–based visual feedback accelerates user learning and improves control in bilateral exoskeleton therapy.⁴ These findings underscore the clinical potential of EMG–driven control strategies for adaptive, patient–specific rehabilitation interventions. However, accurately translating raw EMG signals into meaningful muscle activation inputs requires careful signal preprocessing. The upper limb presents unique challenges due to its complex, multi-joint movements and often weaker or more variable muscle activations (especially in clinical populations). A robust EMG preprocessing pipeline is therefore critical to ensure the model receives high-quality inputs. Key preprocessing steps typically include filtering to remove noise/movement artifacts, extracting a smoothed activation envelope, thresholding to detect true muscle activity, and normalizing signal magnitudes.⁵ Recent studies emphasize that generic EMG preprocessing pipelines (often borrowed from lower-limb applications) may not suffice for upper-limb modeling, motivating refined or individualized approaches.⁶

Upper-limb modeling presents specific challenges not observed in lower-limb simulations. Due to the finer motor control, smaller muscle groups, and more variable recruitment strategies in tasks like elbow flexion or wrist rotation, signal-to-noise ratio (SNR) tends to be lower, and EMG variability higher across subjects and condition.⁷ These factors make generalized EMG filtering and normalization parameters ineffective. For instance, a fixed high-pass cutoff or envelope frequency may over-smooth low-amplitude signals or retain noise in highly active muscles, leading to misrepresented activations.

Currently, there are several MS modeling platforms that can implement EMG driven models of the upper extremities, i.e., BoB, Anybody, OpenSim, etc.^7–9 Generating each of their inputs also requires a slightly different signal preparation procedure. Among existing platforms, OpenSim has become a widely used open-source environment for constructing and analyzing MS models, enabling forward and inverse dynamic simulations based on user-defined control strategies.^3,4 However, the accuracy of EMG-driven simulations heavily depends on the quality and preprocessing of the EMG signals, which are inherently noisy, variable across trials, and sensitive to electrode placement and movement artifacts.¹⁰ Traditional preprocessing pipelines involving fixed bandpass filtering, rectification, and envelope detection often fail to account for inter-muscle differences in signal characteristics or movement-specific variability.^6,7 Detecting low-amplitude EMG signals and quantifying muscle coactivation particularly between agonist–antagonist pairs like biceps and triceps is critical for constructing accurate EMG-driven MS models, yet poses significant technical challenges. Subtle activations are often masked by noise or overshadowed by dominant signals, leading to missed detection and subsequent modeling errors. Conventional amplitude-based methods (e.g., single-threshold detection) frequently fail to capture these nuances, resulting in biases that distort joint-angle estimations. Recent work by Huang et al. (2024) improved onset detection by combining the Hilbert–Huang Transform with marginal spectrum entropy, demonstrating earlier and more reliable identification of low-level muscle activations.¹¹ Likewise, a comprehensive review by Merletti and colleagues (2023) highlighted the effectiveness of dual-threshold and adaptive-threshold algorithms, which dynamically adjust detection criteria to local signal properties and significantly reduce false detections. Additionally, methods inspired by radar systems, such as constant false-alarm rate (CFAR) detection, can adaptively modulate thresholds to maintain consistent sensitivity across different signal levels.¹² Together with coactivation index metrics (e.g., Chen et al., 2024), which measure functional muscle interplay, these advanced techniques offer robust strategies for preserving nuanced muscle behaviors in upper-limb modeling key to enhancing model fidelity in rehabilitation settings.

A standard (traditional) EMG signal preprocessing method refers to a commonly used sequence of signal preprocessing steps historically established in biomechanics and rehabilitation literature, primarily intended to describe general muscle activation rather than to serve as input for numerical or EMG-driven models. This method typically includes: band-pass filtering, full-wave rectification, smoothing and normalization to MVC or task-specific maximum. It is most frequently applied in studies involving healthy individuals and does not account for individual muscle frequency variations or model-specific requirements.

Moreover, methods traditionally used in EMG preprocessing for general neuromuscular studies (e.g., fixed 20–450 Hz band-pass, 50 Hz notch, 6 Hz envelope, RMS smoothing) may not be suitable for control-oriented EMG applications in RS models. These classical pipelines often disregard individual muscle spectral characteristics and are not tailored to movement-phase-specific accuracy.¹⁰ Consequently, rehabilitation-oriented MS models require more refined, muscle-specific strategies. Furthermore, normalization using maximum voluntary contraction (MVC) can be inconsistent across sessions and may not reflect functional engagement during dynamic tasks.¹³ To address these limitations, advanced EMG processing strategies incorporating signal-specific spectral analysis and functional weighting based on muscle contribution to movement have been proposed.¹⁴ Sartori et al. (2014) addressed key challenges in EMG-driven modeling by introducing a hybrid neuromusculoskeletal framework that balances direct EMG-based excitations with static optimization. This approach mitigates the dependency on noisy or unreliable EMG signals and improves the tracking of joint moments. However, their work focused on the lower limb and did not explore how signal processing parameters such as spectral thresholds, envelope filtering, or functional normalization affect upper-limb model outputs.

Recent studies have explored various EMG preprocessing pipelines to improve MS models robustness and accuracy. For instance, Li et al. (2023) developed a muscle synergy–based EMG model for wrist rotation prediction, utilizing synergy decomposition of pre-processed EMG signals. Their preprocessing included band-pass filtering, rectification, and smoothing, followed by dimensionality reduction to identify muscle synergies. The model achieved a Pearson correlation coefficient r > 0.9 and normalized RMSE between 5–10% for wrist tasks.¹⁵

In contrast, Barjavel et al. (2023) implemented a multichannel EMG-assisted model to estimate joint moments during upper limb functional tasks. EMG signals were preprocessed using standard band-pass filtering (20 Hz–450 Hz), RMS smoothing, and amplitude scaling. Their model achieved a joint moment error of ∼3.4% but did not report detailed kinematic error metrics such as RMSE or phase shift.⁸ Similarly, Sarshari et al. (2021) applied EMG-driven simulations for estimating muscle co-contraction using rectified, low-pass filtered EMG inputs. Although the model improved co-contraction estimates, it lacked kinematic validation and phase-specific accuracy assessments.¹⁶

In terms of signal preprocessing strategy, Tahmid et al., (2024) adopted machine learning algorithms to map EMG inputs to muscle activations in elbow tasks. Their preprocessing involved high-pass filtering, full-wave rectification, and normalization, achieving R values up to 0.99 for simple movements and around 0.71 for complex upper-limb actions.¹⁷ However, temporal alignment metrics such as phase shift were not included. It is observed that the period 2019–2025 has seen a shift in focus toward the individualization of EMG preprocessing: researchers increasingly adjust filter cut-offs, window lengths, and normalization references to suit specific muscles, subjects, and clinical conditions. Common trends include using band-pass ranges on the order of 10–500 Hz (with flexibility to go higher on the high-pass if needed), envelope filters around 5–10 Hz (or adaptive windows) for a balance of smoothness and responsiveness, and moving beyond MVC-centric normalization to more practical methods for impaired users.¹⁷ Generic pipelines, when applied blindly to upper-limb data, can lead to significant accuracy issues from mis-detecting muscle onset to mis-scaling activation magnitudes.¹⁸ The literature strongly suggests that by tailoring preprocessing to the individual (and by extension calibrating models on a per-subject basis), we can achieve more reliable simulations of muscle forces and more precise control of assistive devices.¹⁹

The most commonly selected EMG signal preprocessing parameters in the literature^5,8,16,20 are presented in Table 1.

Table 1.
The most commonly used EMG signal preprocessing parameters.

Author Modeling platform/ model Application HPF/LPF Filter order Envelope Normalization Threshold or FFT/ PSD use

Sartori et al., 2014¹⁴ OpenSim + CEINMS Real-time force estimation 20 Hz/6 Hz 4th Yes (6 Hz) MVC No

Ao et al., 2023 ²¹ OpenSim Personalized musculoskeletal trunk model Not specified 4th Yes (6 Hz) MVC Yes (0.10%)

Pizzolato et al., 2015²² OpenSim + CEINMS Gait and upper limb rehab 20 Hz/6 Hz 2nd–4th Yes (6 Hz) MVC No

Laksono et al., 2021²³ AnyBody/OpenSim Two DoF upper-limb robot Unspecified 4th Yes (5 Hz) MVC + synergy factors Yes (0.05%)

Tanaka et al., 2022²⁴ EMG–ML/upper limb Real-Time motion prediction using electromyography 20 Hz/450 Hz 4th Yes (5 Hz) MVC Yes (0.05%)

Lloyd & Besier, 2003²⁵ EMG-to-force model Muscle force estimation 40 Hz/6 Hz Not specified Yes (6 Hz) MVC No

Mathieu et al., 2025²⁶ OpenSim Elbow rehab research 20–460 Hz input HPF (from coherence study) 4th–order band-pass, adaptive per muscle Yes (4 Hz) Not mentioned Yes (0.05%)

Visser & Medendorp, 2023¹¹ General Review Typically HPF 20–450 Hz in cited methods 2nd Yes (6 Hz) MVC PSD threshold = 1–3 SD above baseline

Author	Modeling platform/ model	Application	HPF/LPF	Filter order	Envelope	Normalization	Threshold or FFT/ PSD use
Sartori et al., 2014¹⁴	OpenSim + CEINMS	Real-time force estimation	20 Hz/6 Hz	4th	Yes (6 Hz)	MVC	No
Ao et al., 2023 ²¹	OpenSim	Personalized musculoskeletal trunk model	Not specified	4th	Yes (6 Hz)	MVC	Yes (0.10%)
Pizzolato et al., 2015²²	OpenSim + CEINMS	Gait and upper limb rehab	20 Hz/6 Hz	2nd–4th	Yes (6 Hz)	MVC	No
Laksono et al., 2021²³	AnyBody/OpenSim	Two DoF upper-limb robot	Unspecified	4th	Yes (5 Hz)	MVC + synergy factors	Yes (0.05%)
Tanaka et al., 2022²⁴	EMG–ML/upper limb	Real-Time motion prediction using electromyography	20 Hz/450 Hz	4th	Yes (5 Hz)	MVC	Yes (0.05%)
Lloyd & Besier, 2003²⁵	EMG-to-force model	Muscle force estimation	40 Hz/6 Hz	Not specified	Yes (6 Hz)	MVC	No
Mathieu et al., 2025²⁶	OpenSim	Elbow rehab research	20–460 Hz input HPF (from coherence study)	4th–order band-pass, adaptive per muscle	Yes (4 Hz)	Not mentioned	Yes (0.05%)
Visser & Medendorp, 2023¹¹	General	Review	Typically HPF 20–450 Hz in cited methods	2nd	Yes (6 Hz)	MVC	PSD threshold = 1–3 SD above baseline

*HPF—high-pass filtering; LPF—low-pass filtering; FFT—fast Fourier transformation; PSD—power spectral density.

Although researchers are conducting a lot of research to improve EMG input generation strategies, there are still no consistent algorithms. Therefore, given the current limitations, the aim of this study is to develop and validate a preprocessing framework by performing: 1)

Muscle-specific FFT/PSD-based filtering with variable thresholds,

Envelope smoothing with tuneable frequency cutoffs,

Dual-mode normalization (MVC and functional weights from biomechanical literature),

And comparative validation against motion capture system.

The proposed approach is evaluated by systematically varying key parameters to identify optimal combinations that maximize joint kinematic accuracy. This methodology advances EMG-driven modeling by introducing a physiologically grounded and adaptable preprocessing framework.

It is important to emphasize that the present study does not aim to provide clinical diagnostic recommendations or population-level health status decisions. Instead, the objective is to develop and technically evaluate a personalized EMG-driven musculoskeletal modeling framework under controlled experimental conditions. The investigated cohort represents a homogeneous group of healthy young adults, selected to minimize inter-subject variability and isolate methodological performance. Therefore, the results should be interpreted as a proof-of-concept validation of the modeling strategy rather than a clinically generalizable outcome.

2 Methods

2.1 Participants and exercise

To collect EMG data set, five healthy adults (3 males, 2 females; age 22 ± 1 years) volunteered for the test. Inclusion criteria included the absence of musculoskeletal disorders or prior upper limb injuries. Each subject completed 10 repetitions of elbow flexion–extension exercise (approximately from a 0° to 90° range) using the dominant (right) arm, resulting extension movements in a total of 50 trials (n = 50).

Elbow extension-flexion was performed. The start of the movement and the reference point of the elbow were considered when the arm was fully extended.

2.2 Data acquisition

Surface electromyographic (sEMG) signals were recorded using a Delsys Trigno wireless system (Delsys Inc., USA) at a sampling rate of 2000 Hz. Sensors were placed on six upper limb muscles: biceps brachii (BIClong), triceps brachii (TRIlong), brachialis (BRA), brachioradialis (BRD), flexor carpi radialis (FCR) and extensor carpi radialis (longus) (ECR) (Figure 1). The selection of the muscles was grounded in their primary functional roles during elbow flexion-extension and forearm stabilization, which are central to upper-limb rehabilitation and EMG-driven modeling tasks. These muscles represent a functionally relevant set for capturing the neuromuscular control of elbow motion, especially in dynamic and rehabilitative contexts, as also emphasized in recent EMG-driven modeling studies.^14,17 Their combined inclusion enhances model input fidelity, enabling more accurate prediction of joint kinematics and improved responsiveness to muscle-specific activation profiles.

Figure 1.

EMG electrode placement: 1—biceps brachii; 2—triceps brachii; 3—brachialis; 4—brachioradialis; 5—flexor carpi radialis; 6—extensor carpi radialis.

Upper limb kinematics were recorded using the Xsens Awinda (Xsens Technologies B.V., Netherlands) motion capture system. For this study, the upper limb protocol was employed, with sensors placed on the head, sternum, shoulders, upper arm (humerus), forearm (radius/ulna), dorsal side of the hand of the right arm and pelvis.²⁷ Calibration was performed using the standard “T-pose” procedure.

During the measurement, subjects performed a single cycle of right elbow flexion and extension, beginning from a fully extended position (0°) and flexing to approximately 90°, then returning to the starting posture. Kinematic data were sampled at 60 Hz and exported as .xlsx files using Xsens MVN Analyze software. These sensors record the spatial orientation of each segment through quaternions or Euler angles. Using the internal Xsens biomechanical model, the software calculates joint angles between segments based on their relative orientation. The elbow joint angle during flexion/extension movement was used in the study.

The Xsens system was synchronized with surface EMG recordings using an external trigger to ensure temporal alignment between muscle activity and joint kinematics. This enabled accurate comparison between EMG-driven OpenSim outputs and Xsens-derived motion trajectories.

2.3 EMG signal preprocessing

To generate physiologically reliable inputs for an EMG-driven MS model, a multi-stage EMG preprocessing pipeline was employed. Each step was carefully selected based on prior validation studies^8,10,16,17 and literature consensus, with additional comparative analyses conducted in this study to determine the most optimal parameters.

In this study EMG data were processed offline using custom MATLAB (MathWorks, MA) scripts. In order to find out the most appropriate preprocessing strategy of the EMG signal, the signal of the muscles was preprocessed. Before the main preprocessing steps, which are presented in Figure 2, the activity of all six muscles was cropped from the beginning to the end of the elbow flexion/extension movement, then the signal was detrended, and the time was normalized to the flexion/extension movement cycle (%).

Figure 2.

EMG signal preprocessing strategy with functional weighting normalization.

1. Spectral Analysis (FFT and PSD)

The objective of this step was to identify muscle-specific signal characteristics and frequency content. Muscle signals were analyzed using fast Fourier transform (FFT) and Welch's power spectral density (PSD) to determine dominant frequencies and signal-to-noise ratios. Three spectral thresholds—0.01%, 0.05%, and 0.1% of cumulative power were applied to define effective bandwidths. These threshold values were selected to assess the impact of different levels due to the following reasons: 0.01% is conservative and retains low-amplitude but physiologically meaningful activations, such as antagonist co-activation, making it suitable for sensitive rehabilitation models; 0.05% provides a balance between noise reduction and signal fidelity and is commonly used in upper-limb musculoskeletal simulations²⁸; 0.1% excludes a larger portion of the power spectrum, which may be beneficial in noisy recordings but can reduce accuracy in tasks involving fine motor control.²⁹

2. Adaptive Filtering

To account for muscle-specific variability and ensure the preservation of physiologically relevant signal components, an adaptive filtering strategy was employed. Rather than applying uniform filter cutoffs across all muscles, the high-pass and low-pass filter parameters were determined individually for each muscle based on its spectral analysis.

For each EMG channel, the algorithm calculated the normalized cumulative power from the FFT/PSD curve and identified the frequency at which a predetermined threshold (e.g., 0.01%, 0.05%, or 0.1% of total spectral power) was reached. This threshold-guided approach ensured that the majority of meaningful spectral content was preserved while minimizing noise and motion artefacts.

In addition to selecting individual cutoff frequencies, the effect of different filter orders was also evaluated. Different Butterworth filter orders (2nd, 4th, and 5th) were tested based on the most commonly used parameters reported in the reviewed literature (Table 1).

3. Rectification

Full-wave rectification was applied to convert the bipolar signal into an absolute-valued series, allowing for subsequent envelope computation. This step retains the overall temporal pattern of muscle activation while simplifying amplitude processing.

4. Envelope Smoothing

The signals were smoothed using the 4th order low-pass Butterworth filter to generate a linear envelope. The cut-off frequencies used for envelope generation were 3 Hz, 4 Hz, and 5 Hz, based on the most commonly used envelope frequencies in upper limb musculoskeletal models.^8,30 3 Hz is considered suitable for capturing slow-varying movements and minimizing phase distortion, 4 Hz is often used in rehabilitation-focused simulations due to its balance between smoothness and temporal responsiveness and 5 Hz is recommended when higher temporal resolution is needed, particularly in dynamic elbow flexion-extension tasks.

5. Normalization

Two normalization strategies were implemented: 1)

MVC-based normalization—each muscle was normalized to its maximum voluntary contraction (MVC).

Functional weighting—EMG amplitudes were further scaled using phase-specific activation weights derived from literature, reflecting each muscle's role in elbow flexion and extension.

5.1. MVC-based min–max normalization

To standardize EMG amplitude across muscles and trials, each processed envelope was scaled using min–max normalization. The maximum value was derived from the highest dynamic peak observed across all trials and muscles for each participant, representing 100% activation. This approach was adopted in place of static maximum voluntary contraction (MVC) trials, which can be inconsistent and may underestimate peak activation during dynamic tasks.³¹ The resulting normalized signals ranged from 0 to 1 and allowed inter-muscle and inter-trial comparisons on a unified scale.⁶

5.2. Functional weighting normalization

Following amplitude normalization, each muscle's signal was further scaled using task-specific biomechanical weights that reflected its contribution to elbow torque production during flexion or extension. These weights were derived from literature-reported estimates of physiological roles and anatomical leverage (Table 2):

Table 2.

Muscle weights during elbow flexion (Buchanan et al., 2004; Holzbaur et al.,2005³²).

Muscle	Weight (%)
Biceps brachii	35
Triceps brachii	10
Brachioradialis	15
Extensor digitorum	5
Flexor Carpi Radialis	5

These weights were established based on inverse dynamics simulations, electromyographic patterns, and muscle moment arm analyses reported in previous biomechanical studies.^32,33 Functional muscle weights, which describe the relative contribution of each muscle to a specific movement, were determined using computational modeling and biomechanical analysis methods. Buchanan et al. (2004) reviewed EMG-driven musculoskeletal models in which muscle forces and their contributions to joint moments were calculated using inverse dynamics and optimization algorithms, based on muscle activation signals, anatomical structure, and mechanical properties. A similar approach was applied by Holzbaur et al. (2005), who developed an upper limb musculoskeletal model based on MRI data and anatomical atlases. In this model, each muscle's contribution to movement was evaluated according to its physiological cross-sectional area (PCSA), moment arm, and estimated activation during the motion. Functional weights were derived as the proportions of simulated muscle forces during a specific movement such as elbow flexion or extension. It is important to note that these functional weights are not fixed and may vary depending on the specific task (e.g., unloaded vs. loaded movement, movement speed, joint angle range), population characteristics (age, sex, pathology), and biomechanical constraints. Since the subjects are healthy adults and the investigated movement is isolated and biomechanically well-defined (elbow flexion/extension), the functional weights defined in the literature are sufficiently generalizable and can be applied in this study as a justified normalization approach. Moreover, this allows for a more accurate evaluation of antagonist and synergist muscle involvement and helps optimize EMG input for the EMG-driven musculoskeletal model in the OpenSim environment.

Muscles like the brachialis were accounted for through their synergistic contribution with the biceps and brachioradialis.^34,35 In some cases, phase-specific adjustments were applied. For example, the triceps brachii signal was amplified in the final 20% of the movement cycle to compensate for known underactivation in terminal elbow extension due to low EMG signal amplitude.³⁶ This ensured sufficient torque generation during the extension phase.

2.4 Determination of preprocessing strategy

The EMG preprocessing strategy consisted of four main steps: selection of filter order, determination of filter cutoff frequencies, envelope extraction, and normalization. First, the strategies were tested by changing one of the parameters of these steps. This was done to evaluate the influence of each parameter on the output of the musculoskeletal (MS) model. The tested strategy variants are presented in Table 3.

Table 3.
Preprocessing strategy variants.

Variant Filter order Threshold, % Envelope, Hz Normalization

C1-1 2nd 0.05 3 MVC + functional weights

C1-2 4th

C1-3 5th

C2-1 4th 0.01 3 MVC + functional weights

C2-2 0.05

C2-3 0.1

C3-1 4th 0.05 3 MVC + functional weights

C3-2 4

C3-3 5

C4 4th 0.05 3 MVC only

C5 4th 0.05 3 MVC + functional weights

Variant	Filter order	Threshold, %	Envelope, Hz	Normalization
C1-1	2nd	0.05	3	MVC + functional weights
C1-2	4th
C1-3	5th
C2-1	4th	0.01	3	MVC + functional weights
C2-2	0.05
C2-3	0.1
C3-1	4th	0.05	3	MVC + functional weights
C3-2	4
C3-3	5
C4	4th	0.05	3	MVC only
C5	4th	0.05	3	MVC + functional weights

Parameter combinations were also tested, changing not one but several different parameters. The purpose of creating combinations was to investigate the influence (compensation) of different parameters on the MS model simulation. The applied combinations are presented in Table 4.

Table 4.

Parameter combinations.

Combination	Filter order	Threshold, %	Envelope, hz	Normalization
C6	5th	0.01	3	MVC + functional weights
C7	4th	0.05	5	MVC + functional weights
C8	4th	0.1	3	MVC + functional weights
C9	2th	0.05	3	MVC + functional weights

Each combination focused on modifying a specific signal processing component while keeping others constant to isolate its effect: 1)

Combination 1 (C6): Applied a 0.01% FFT/PSD threshold with a 5th order Butterworth filter, aiming to test the effect of a stricter spectral cutoff on activation clarity.

Combination 2 (C7): Used a 0.05% threshold and a 5 Hz low-pass envelope, selected based on literature for optimal temporal smoothing and cutoff precision.

Combination 3 (C8): Implemented a 0.1% threshold with a 3 Hz envelope, to observe how a lenient spectral cutoff combined with a slower envelope affects timing and amplitude.

Combination 4 (C9): Applied a 2nd order Butterworth filter and a combined MVC and functional weight normalization, focusing on how filter sharpness and physiological weighting together impact model output.

2.5 Musculoskeletal model and forward simulation

A modified version of the upper-limb MS model developed by Holzbaur et al. (2005) was implemented in OpenSim (Figure 3). The model includes anatomically accurate representations of bones, joints, and musculotendon actuators for the upper extremity.³² For this study, the shoulder joint was locked, and only elbow flexion–extension was simulated as a single degree of freedom (DOF) movement.

Figure 3.

A modified Holzbaur et al., 2005³² version of the OpenSim upper-limb model: (a) initial position of the movement; (b) final position of the movement.

2.6 Model inputs and configuration

Each of the EMG preprocessing strategies was used as input to the OpenSim musculoskeletal (MS) model. The six normalized and functionally weighted EMG signals were mapped to their respective muscles in the model and used as direct excitation inputs for forward dynamics simulation. No reserve actuators were included in the simulation, all joint motion was driven entirely by EMG-based muscle excitations. Muscle-tendon dynamics, including activation–contraction coupling and force–length–velocity relationships, were retained as defined in the original OpenSim model.

Each of the 50 movement trials was simulated independently using OpenSim Forward Dynamics Tool. Simulated joint angles were exported for comparison with experimentally recorded joint kinematics.

2.7 Validation and statistical analysis

The performance of the EMG-driven MS model was assessed by comparing simulated elbow joint angles to reference kinematic data recorded via the Xsens motion capture system. For each trial, the elbow flexion–extension trajectory generated by OpenSim was time-normalized to a standardized 0–100% movement cycle and interpolated to match the temporal resolution of the Xsens data.

All statistical analyses were conducted in MATLAB software.

Model accuracy was evaluated using the following statistical measures:

Root mean square error (RMSE) to assess average deviation across the full movement cycle.

Mean absolute error (MAE) to capture absolute differences regardless of direction.

Pearson correlation coefficient (r) to evaluate the strength of linear agreement between simulated and recorded angles.

Phase shift to measure temporal alignment by calculating the time lag of the peak elbow flexion angle.

The Bland–Altman method was used to evaluate the level of agreement between the EMG-driven model and the Xsens reference system. This involved calculating the mean bias (systematic error) and 95% limits of agreement (LoA), defined as ±1.96 standard deviations around the mean difference. Visual inspection of Bland–Altman plots was used to check for heteroscedasticity or trend effects.

3 Results

The results of this study are presented sequentially according to the preprocessing stages applied to the EMG signals. Each section evaluates the influence of a specific parameter or preprocessing step beginning with spectral thresholding (FFT/PSD), followed by envelope frequency selection, filter order variations, and normalization strategies on the accuracy of the EMG-driven musculoskeletal model. Quantitative comparisons are made against reference kinematic data (Xsens), and metrics such as RMSE, MAE, correlation coefficient (r), and phase shift are used to assess performance at each stage. The goal is to systematically identify the parameter configurations that most effectively enhance the model's alignment with experimentally measured joint angles.

3.1 Effect of filter order on model accuracy

As shown in Figure 4(a–(c)), significant differences were observed between filter configurations.

Figure 4.

The influence of different filter orders on the output of an OpenSim MS model driven by EMG signals.

The 2nd order filter resulted in the poorest performance, with RMSE of 32.42°, MAE of 29.15°, a low correlation coefficient (r = 0.610), and a substantial phase shift (+10.5%). This low order was insufficient to attenuate high-frequency noise, resulting in amplified signal variability and substantial deviation from the reference Xsens trajectory.

In contrast, the 4th order filter demonstrated the highest accuracy, yielding RMSE = 8.52°, MAE = 7.26°, a strong correlation (r = 0.986), and minimal phase shift (−2.9%). These values suggest optimal smoothing with adequate preservation of signal shape and timing, aligning well with the gold-standard kinematics.

The 5th order filter produced intermediate results (RMSE = 17.32°, MAE = 13.23°, r = 0.841, phase shift −1.8%). Although the phase agreement remained acceptable, the curve exhibited signs of over-smoothing, especially during the motion deceleration phase (60–100%), indicating loss of dynamic fidelity.

The 4th order filter provided the best trade-off between noise suppression and temporal fidelity, consistent with recent literature recommendations for EMG envelope extraction in upper-limb models used for rehabilitation purposes.^8,17

3.2 Elbow joint angles comparison using differences between FFT/PSD parameters

Three FFT/PSD based threshold values (0.01%, 0.05%, and 0.1%) were applied to define the adaptive band-pass filters used in EMG preprocessing (Figure 5). Figure sets show each comparison with average curves and standard deviation (±SD) envelopes.

Figure 5.

Comparison of Xsens and OpenSim elbow angle due to the threshold: (a) 0.01%; (b) 0.05%; (c) 0.1%.

Selected 0.01% threshold allowed a wide frequency range to pass, including very low amplitude components, which may include noise. As a result, the EMG signals may be overly sensitive to minor signal fluctuations, leading to early activation artifacts and overestimation of motion amplitude (peak ∼20–30° above Xsens reference). Despite a decent correlation, temporal misalignment and inflated amplitude reduce the overall accuracy.

Threshold 0.05%—this intermediate value provided the best trade-off between sensitivity and noise rejection. The resulting elbow angle trajectory closely follows the Xsens reference throughout the motion cycle, with minimal amplitude inflation and almost no temporal drift. It also yields the narrowest SD envelope, indicating consistent inter-trial behavior. Thus, the 0.05% threshold appears to be the most suitable value for defining frequency limits in adaptive filtering for this dataset.

Using 0.1% threshold the correlation result is numerically high and the actual accuracy is poor. This threshold filters out too many lower amplitude components, leading to underrepresentation of subtle muscle activations. Consequently, the model consistently overestimates joint angles due to a lack of nuanced EMG modulation, particularly during the transition phases. The standard deviation is wide, indicating poor consistency.

Prior studies support the need for carefully balanced filtering. Hodson-Tole & Wakeling (2009) emphasize the need to retain subject-specific and muscle-specific frequency content to maintain physiological signal integrity. Similarly, Farina et al. (2014) argue that over-aggressive filtering may distort onset timing and motor unit recruitment signals.^14,15

Using fixed band-pass filtering (e.g., 20–450 Hz) may be overly generic. The adaptive approach used here allows signal-specific cutoff determination, and results show that 0.05% energy threshold yields optimal alignment between simulated and reference motion, making it ideal for applications like EMG-driven rehabilitation modeling.

3.3 Impact of envelope filter frequency on model accuracy

Figure 6 illustrates the impact of different envelope cutoff frequencies (3 Hz, 4 Hz, and 5 Hz) on the agreement between OpenSim and Xsens elbow flexion/extension angles across the motion cycle.

Figure 6.

Comparison of Xsens and OpenSim elbow angle: (a) 3 Hz envelope frequency; (b) 4 Hz envelope frequency; (c) 5 Hz envelope frequency.

The signal appears overly smoothed, leading to reduced responsiveness and a flattened movement curve. This caused a notable temporal shift in the trajectory (phase shift: −14.4%) and a moderate correlation (r = 0.874). The high RMSE (17.87°) and MAE (13.01°) indicate a loss of physiological fidelity.

A slight improvement in correlation (r = 0.972) and timing (phase shift: −4.8%) was observed, though amplitude errors (MAE: 17.43°) remained relatively high. While more reactive than 3 Hz, the smoothing was still suboptimal for dynamic joint tracking.

The 5 Hz envelope provided the most physiologically consistent and temporally accurate joint trajectory. The curve demonstrated high similarity to the Xsens reference both in shape and amplitude. The correlation was extremely strong (r = 0.986), and the minimal phase shift (−1.1%) confirms temporal alignment. This frequency is also supported in literature for upper-limb applications as a good balance between noise suppression and signal fidelity.²

The findings align with established EMG preprocessing guidelines, which recommend envelope low-pass filters in the 3–6 Hz range for upper-limb dynamic tasks¹⁰ Frequencies above 5 Hz tend to pass motor unit firing variability and electrical noise, deteriorating the stability of muscle activation profiles.

The superior results with the 5 Hz envelope validate its appropriateness for real-time or predictive EMG-driven modeling in rehabilitation, providing both accuracy and physiological plausibility.

3.4 Comparison using MVC normalization

Figure 7 presents the comparison between the Xsens elbow flexion-extension angles and the OpenSim output driven by EMG signals normalized using only maximum voluntary contraction (MVC). The results reveal substantial deviations between the two curves throughout the motion cycle. The OpenSim derived elbow angle demonstrates a significantly higher variability and reduced alignment with the reference Xsens signal. Quantitatively, this is reflected by a high root mean square error (RMSE) of 43.24°, a mean absolute error (MAE) of 27.35°, a weak Pearson correlation coefficient (r = 0.285), and a noticeable phase shift of +6.3%. These results suggest that global MVC normalization alone may lack the precision and contextual adaptability required to accurately reflect dynamic muscle contributions during functional elbow flexion tasks.

Figure 7.

Comparison of Xsens and OpenSim elbow angle using a MVC normalization only.

3.5 Final output comparison

Figure 8, the model achieved a root mean square error (RMSE) of 7.78°, a mean absolute error (MAE) of 6.72°, a Pearson correlation coefficient of r = 0.980, and a phase shift of 2.6%. The shape and timing of the predicted joint angle trajectory closely mirrored the experimental data, with consistent overlap throughout the motion cycle. This result validates the efficacy of the personalized and functionally-informed EMG preprocessing approach in producing physiologically accurate kinematic predictions suitable for rehabilitation and biomechanical modeling applications.

Figure 8.

Comparison of Xsens and EMG driven OpenSim MS model elbow angles using combined normalization.

The average elbow angle trajectory from Xsense exhibited a smooth sinusoidal pattern with a peak flexion (∼80°) occurring around the midpoint (50%) of the movement cycle.

The OpenSim output closely followed this trend, displaying strong temporal alignment and minimal amplitude offset. A minor phase lead but demonstrated a slight phase lead (∼2–3%) was observed in the EMG-driven output, along with slight underestimate during terminal extension.

3.6 Comparison of parameter combinations for EMG-driven joint angle prediction

Figure 9(a-(d)) presents the averaged elbow flexion/extension curves with corresponding standard deviations across all participants.

Figure 9.

Influence of C6-C9 combinations on MS model output: (a) combination 1, (b) combination 2, (c) combination 3 and (d) combination 4.

Configuration 1 (C6) (0.01% threshold, 5th-order Butterworth) produced a moderately accurate output with RMSE of 10.64°, MAE of 8.33°, correlation coefficient r = 0.920, and a phase shift of −10.8%. Although the fifth-order filter improved noise attenuation, the stricter frequency threshold (0.01%) may have removed useful signal components, resulting in noticeable amplitude underestimation during the peak flexion phase.

Configuration 2 (C7) (0.05% threshold, 5 Hz envelope) demonstrated the best overall agreement with the reference: RMSE = 9.86°, MAE = 7.02°, r = 0.949, and phase shift = −8.5%. The 0.05% threshold preserved relevant muscle activity while excluding high-frequency noise, and the 5 Hz envelope maintained physiological envelope smoothness. These results support this configuration as the optimal trade-off between accuracy and generalizability.

Configuration 3 (C8) (0.1% threshold, 3 Hz envelope). The loosest thresholding (0.1%) combined with a low envelope frequency led to an RMSE of 10.93°, MAE of 8.07°, r = 0.926, and phase shift of −10.0%. Despite the acceptable RMSE and correlation, the resulting signal exhibited early saturation and over-smoothed motion dynamics, likely due to the combination of low-pass filtering and high spectral inclusion.

Configuration 4 (C9) (2nd-order Butterworth, combined MVC + functional weight normalization) balanced simplicity and robustness, yielding an RMSE of 11.14°, MAE of 9.01°, r = 0.967, and a phase shift of −8.3%. While slightly less accurate than higher-order filtering cases in RMSE, the combined normalization appears to enhance temporal fidelity and amplitude consistency, aligning well with rehabilitation-focused outcomes.

Overall, the comparative results emphasize that both overly restrictive (0.01% threshold) and overly permissive (0.1% threshold) spectral cutoffs degrade signal quality. Furthermore, the 5 Hz envelope frequency provided a more reliable envelope reconstruction compared to 3 Hz. Notably, combined normalization (MVC + muscle-specific weights) consistently improved signal scaling and phase alignment.

3.7 Quantitative accuracy metrics

The statistical comparison between the EMG-driven simulations and motion capture data demonstrated strong agreement (Table 5).

Table 5.
Error metrics summary (RMSE, MAE, r, phase shift, statistical significance (p-value)).

Variant Marking RMSE (°) MAE (°) r Phase shift (%) p-value*

2nd order Butterworth C1-1 32.42 29.15 0.610 +10.5 <0.01

4th order Butterworth C1-2 8.52 7.26 0.986 −2.9 <0.01

5th order Butterworth C1-3 17.32 13.23 0.841 −1.8 <0.01

FFT/PSD 0.01% C2-1 15.80 11.80 0.901 −12.8 0.87

FFT/PSD 0.05% C2-2 14.84 11.32 0.937 +2.3 <0.01

FFT/PSD 0.1% C2-3 23.40 21.87 0.967 −5.9 <0.01

Envelope 3 Hz C3-1 13.87 13.01 0.874 −14.4 <0.01

Envelope 4 Hz C3-2 18.90 17.43 0.972 −4.8 0.62

Envelope 5 Hz C3-3 10.28 9.28 0.986 −1.1 <0.01

MVC normalization C4 43.24 27.35 0.285 +6.3 <0.01

MVC + functional weights C5 7.78 6.72 0.980 +2.6 0.87

C1 combination C6 10.64 8.33 0.920 −10.8 0.87

C2 combination C7 9.86 7.02 0.949 −8.5 <0.01

C3 combination C8 10.93 8.07 0.926 −10.0 <0.01

C4 combination C9 11.14 9.01 0.967 −8.3 0.02

Variant	Marking	RMSE (°)	MAE (°)	r	Phase shift (%)	p-value*
2nd order Butterworth	C1-1	32.42	29.15	0.610	+10.5	<0.01
4th order Butterworth	C1-2	8.52	7.26	0.986	−2.9	<0.01
5th order Butterworth	C1-3	17.32	13.23	0.841	−1.8	<0.01
FFT/PSD 0.01%	C2-1	15.80	11.80	0.901	−12.8	0.87
FFT/PSD 0.05%	C2-2	14.84	11.32	0.937	+2.3	<0.01
FFT/PSD 0.1%	C2-3	23.40	21.87	0.967	−5.9	<0.01
Envelope 3 Hz	C3-1	13.87	13.01	0.874	−14.4	<0.01
Envelope 4 Hz	C3-2	18.90	17.43	0.972	−4.8	0.62
Envelope 5 Hz	C3-3	10.28	9.28	0.986	−1.1	<0.01
MVC normalization	C4	43.24	27.35	0.285	+6.3	<0.01
MVC + functional weights	C5	7.78	6.72	0.980	+2.6	0.87
C1 combination	C6	10.64	8.33	0.920	−10.8	0.87
C2 combination	C7	9.86	7.02	0.949	−8.5	<0.01
C3 combination	C8	10.93	8.07	0.926	−10.0	<0.01
C4 combination	C9	11.14	9.01	0.967	−8.3	0.02

≥0.80—excellent agreement (no significant difference);

0.50-0.79—good agreement;

0.05–0.49—possible difference;

<0.05—significant difference (poor agreement)

*Higher p-value indicate better agreement between OpenSim and Xsens joint angles.

According to established criteria for model validity (e.g., RMSE < 8°, r > 0.95), the EMG-driven simulation output is considered both accurate and reliable for reproducing physiological elbow motion in this task context.

Based on the summarized results across all tested preprocessing variations, it is evident that the EMG-driven MS model achieved the highest accuracy when a combination of FFT/PSD threshold at 0.05%, a 5 Hz envelope smoothing frequency, and normalization based on MVC with functional muscle weighting was applied. This configuration yielded the lowest RMSE (7.78°) and MAE (6.72°), along with a high correlation coefficient (r = 0.980) and minimal phase shift (2.6%). These metrics demonstrate not only improved amplitude tracking but also enhanced temporal alignment of the simulated joint trajectory compared to the Xsens reference. In contrast, alternative configurations particularly those employing higher envelope frequencies or standalone MVC normalization resulted in substantially larger errors and phase discrepancies. This emphasizes the importance of personalized preprocessing and dual-stage normalization, aligning with findings from recent studies but extending them by demonstrating that fine-tuned preprocessing choices significantly affect model fidelity.

A comparison of the results of the basic EMG signal processing and the optimal configuration is presented in Table 6.

Table 6.

Comparison of baseline (static MVC-only) and optimal EMG signal processing configurations

Variant	RMSE (°)	MAE (°)	r	Phase shift (%)	p-value
Optimal configuration	7.78	6.72	0.980	+2.6	0.87
Baseline (MVC only)	43.24	27.35	0.285	+6.3	< 0.01

These results validate the proposed pipeline as a robust and physiologically grounded approach for EMG-based joint kinematics estimation in upper-limb rehabilitation contexts. Results also confirm that the adaptive EMG preprocessing and dual-stage normalization strategy effectively reproduces physiological elbow kinematics with high fidelity.

3.8 Bland–altman agreement analysis

The Bland–Altman plot (Figure 10) demonstrated a consistent bias of +7.30° indicating that EMG-driven model slightly overestimated elbow angles compared with Xsense. However, the majority of data points fell with 95% limits of agreement (−4.98° to +19.57°) and no pattern of heteroscedasticity was observed.

Figure 10.

Bland–Altman plot comparing elbow flexion/extension angles from Xsens and OpenSim outputs.

The slight positive bias may be attributed to triceps underactivation during terminal extension or to modeling simplifications such as the lack of neuromuscular delay compensation.

Figure 11 was created encompassing all evaluated variants, including filter threshold levels, envelope cut-off frequencies, Butterworth filter orders, and normalization approaches. The diagram displays four key performance metrics: root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (r), and phase shift.

Figure 11.

Radar chart comparing all tested EMG signal processing configurations based on RMSE, MAE, correlation (r), and phase shift metrics in predicting elbow joint kinematics using an EMG-driven MS model.

From the radar plot, it is evident that the MVC + functional weights configuration demonstrated the most favorable overall performance, characterized by the lowest RMSE (7.78°), lowest MAE (6.72°), highest correlation (r = 0.980), and a relatively minor phase shift (+2.6%). This indicates that combining physiological normalization with task-specific weighting provides a robust modeling strategy in EMG-driven simulations of elbow motion.

Among the combinations (C6–C9), Combination 2 (C7) which applies a 0.05% FFT/PSD threshold and a 5 Hz envelope also performed well, with a strong correlation (r = 0.949) and balanced error metrics (RMSE = 9.86°, MAE = 7.02°), validating the synergistic benefit of optimized filtering and smoothing parameters.

In contrast, configurations such as MVC-only normalization and the 2nd order Butterworth filter presented significantly larger RMSE and MAE values, accompanied by poor correlation and substantial phase shifts. This reinforces the limitations of relying solely on traditional signal processing techniques in upper-limb MS modeling, particularly when inter-muscle coordination and timing accuracy are critical.

The radar diagram thus provides a holistic view of how parameter selection directly influences output accuracy, highlighting the need for multifactorial optimization in EMG signal processing pipelines to enhance MS model fidelity.

4 Discussion

The Discussion aims to clarify the methodological contributions of the proposed EMG-driven framework, interpret the biomechanical findings in the context of existing literature, and outline the potential clinical and translational relevance within the defined modeling scope. Particular attention is given to clearly distinguishing technical validation outcomes from broader generalization claims, ensuring that conclusions remain aligned with the methodological boundaries of the study.

Upper-limb EMG-driven models universally apply filtering, rectification, and smoothing to derive muscle activation signals, but most studies use fixed cutoff frequencies and standard filters. In contrast, the user's method introduces personalized spectral analysis (FFT/PSD) to choose filter parameters adaptively, and applies functional normalization weights, which is uncommon in literature. This could reduce noise and phase delays relative to the fixed pipelines (improving responsiveness crucial in real-time rehab applications and possibly prediction accuracy). However, as seen in these studies, factors like model calibration (e.g., tuning EMG-to-force scaling) and incorporating muscle synergies can also significantly impact performance. The choice of EMG preprocessing methodology influences model output smoothness and timing: for instance, a 4–6 Hz low-pass adds a small lag and smooths out high-frequency noise, whereas an adaptively chosen cutoff might preserve more subject-specific signal details (potentially capturing quicker muscle contractions or unique frequency content). In rehabilitation contexts, performance is typically reported as how well the model's predicted quantities (joint kinematics or kinetics) match measured values (high correlations ∼0.9, low errors ∼10% MVC or < ∼10°), and ultimately how this translates to functional outcomes. The differences in EMG preprocessing pipelines above sometimes yield marginal performance differences (e.g., a few percentage points in error or R $\land$ 2), but robustness and applicability in rehab can be affected by these methods a more personalized filtering approach (like the user's) may improve consistency across patients by accounting for individual signal characteristics, whereas overly generic preprocessing might require more intensive calibration or be less effective for certain patients. Each of these studies contributes evidence that careful EMG preprocessing and model tuning can enhance the predictive power and usability of EMG-driven MS models in rehabilitation settings.

This study demonstrated that adaptive EMG signal preprocessing combined with dual-stage normalization can accurately drive a musculoskeletal simulation of elbow motion in OpenSim. The proposed approach achieved high agreement with motion capture data, with correlation coefficients exceeding 0.98 and joint angle errors remaining below 8°, validating its effectiveness.

The RMSE value (7.78°) are within the range reported in prior EMG-driven modeling studies of the upper limb,^1,13,14 and the phase lead of 2.6% observed in the EMG model output is consistent with previous findings that EMG-driven forward dynamics may respond more rapidly than kinematic recordings due to anticipatory muscle activation.

Unlike conventional EMG pipelines with fixed bandpass cutoffs, the proposed method employed individualized spectral analysis (FFT/PSD) to tailor high-pass and low-pass filters per muscle. This adaptive strategy allowed for retention of meaningful signal bandwidth while reducing subject- and muscle-specific noise. The result was cleaner and more physiologically relevant excitation signals, which translated into more accurate joint simulations. Such tailoring is especially critical in multi-muscle recordings, where frequency components can vary substantially between superficial and deep muscles.³⁷

Moreover, the dynamic MVC-based normalization, derived from the highest peak across all dynamic trials, avoided the limitations of static MVCs, which are often inconsistent or underestimated in dynamic tasks.³⁵ Combined with literature-informed functional muscle weighting, the normalized EMG signals better reflected each muscle's biomechanical role, particularly in the transition between flexion and extension. The application of dynamic triceps weighting during the final 20% of the movement cycle successfully addressed under activation in terminal extension, a common issue in EMG-driven simulations.³⁴

The two-stage normalization process combining dynamic min–max scaling with functional weighting addressed well-known limitations of EMG-based modeling. Dynamic MVC normalization provided a practical alternative to static MVC tests, which often underestimate true activation during dynamic tasks.

Functional weighting ensured that EMG amplitude reflected not just signal magnitude but also each muscle's mechanical role in generating joint torque. This adjustment was especially useful in accounting for underactive signals (e.g., triceps) and ensuring more balanced joint output.

Phase-dependent adjustments (e.g., triceps enhancement during terminal extension) further helped reduce extension underestimation, a common issue in EMG-driven models.

Compared to recent literature, the proposed model performs favorably. For example, Barjavel et al. (2023) reported a mean joint moment prediction error of approximately 3.4% using an EMG-assisted upper-limb model with 16 channels, yet did not report specific kinematic agreement indices such as RMSE or phase shift. Li et al. (2023), who applied a synergy-based EMG model to wrist rotation tasks, reported high correlation values (r > 0.9) and normalized RMSE in the range of 5–10%, which is comparable to the present study's accuracy metrics. Tahmid et al. (2024) reported muscle activation prediction accuracies of r ∼0.99 for simple elbow tasks and r ∼0.71 for complex upper-limb movements, indicating strong internal model performance; however, direct motion tracking accuracy was not their primary outcome.

Importantly, the our present model's phase alignment and correlation metrics (phase shift of +2.6%, r = 0.98) demonstrate stronger temporal precision than most inverse-dynamics-based methods, which are typically less synchronized with the user's neural intent due to the reliance on post-hoc optimization. Moreover, the inclusion of functional muscle weighting and personalized spectral-based filtering distinguishes the current model from many previous studies, which either used generic filtering parameters or lacked functional normalization altogether

From a clinical perspective, the Bland–Altman analysis confirms that the model's prediction error is bounded and systematic, rather than random, making it more suitable for rehabilitation contexts where consistent under- or over-estimation can be compensated algorithmically. Compared to Sarshari et al. (2021), whose EMG-assisted model showed improved co-contraction estimation but lacked phase-specific validation, the current study provides a more comprehensive evaluation of kinematic accuracy in time and amplitude domains.

Overall, the results suggest that the present EMG-driven model not only meets but in some respects exceeds the accuracy and applicability of other state-of-the-art models in the field, particularly for real-time or phase-specific elbow motion tracking in rehabilitation settings. The added benefit of model personalization (both in signal preprocessing and muscle weighting) enhances the model's robustness and translational potential for individualized rehabilitation therapies.

The current validation was performed on a limited cohort of five healthy adults (22 ± 1 years), which ensures internal methodological consistency but restricts external generalizability. The homogeneous age range was intentionally selected to reduce biological variability and allow controlled evaluation of the EMG preprocessing and normalization strategy. However, the behavior of the proposed model in larger datasets, different age groups, or clinical populations remains to be investigated in future studies.

4.1 Clinical relevance and applicability

The optimized EMG signal processing strategy proposed in this study demonstrates strong potential for application in upper-limb rehabilitation and assistive movement modeling. Such precision is critical in clinical settings where joint kinematics must be tracked reliably for therapy progression, prosthetic control, or motor recovery monitoring. Unlike traditional MVC-based normalization, the use of functionally weighted normalization reflects the physiological contribution of muscles during natural tasks, making the approach more ecologically valid. Furthermore, the parametric analysis of filtering and envelope extraction enhances model robustness and adaptability, enabling clinicians and engineers to tailor the processing pipeline to diverse patient conditions. This makes the proposed method especially valuable for neurorehabilitation, EMG-driven exoskeletons, and personalized biomechanical assessments in real-world therapeutic scenarios.

4.2 Limitations and future directions

Despite encouraging results, this study has several limitations. First, the functional weighting factors were derived from literature values rather than subject-specific measurements. While this improves generalizability, inter-individual variability in muscle contribution may impact simulation accuracy.

Second, dynamic MVC-based normalization relies on the assumption that the recorded trials captured maximal voluntary effort. While effortful tasks were used, peak values may still underestimate true maximum activation for some muscles

Third, the use of surface EMG limits the ability to capture deep muscle activity (e.g., brachialis) and introduces susceptibility to signal attenuation and cross-talk. Although functional weighting partially compensates for these limitations, the accuracy of force estimation may still be affected.

Specifically, we acknowledge that including only five healthy individuals limits the generalizability of our findings. Despite the homogeneous nature of the cohort, EMG-driven models are highly sensitive to individual-specific factors such as maximum voluntary contraction (MVC) levels, muscle activation strategies, and anatomical variability. Therefore, our pipeline, though effective within this controlled group, requires subject-specific calibration and cannot yet be assumed to generalize across broader populations without further validation.

The sample size was limited to five healthy young adults, which restricts statistical generalization. The narrow age range does not allow extrapolation of findings to pediatric, elderly, or pathological populations. Also, the study design focuses on methodological validation within a controlled biomechanical context rather than clinical diagnosis or disease classification. According that, the results should be interpreted as evidence of technical feasibility and modeling sensitivity rather than as population-level conclusions.

Finally, the model was restricted to a single degree of freedom elbow flexion–extension and did not account for multi-joint interactions, co-contractions, or neuromuscular delays.

Future work will focus on extending this framework to multi-joint movements, integrating more muscles and degrees of freedom, and incorporating machine learning approaches to estimate muscle excitation patterns. Additionally, subject-specific calibration of functional weights and dynamic modeling of electromechanical delay could further improve simulation realism and generalizability. Also, the future work will include validation on larger and more heterogeneous cohorts, including individuals with neuromuscular impairments, to evaluate robustness and translational applicability.

5 Conclusions

This study presented a validated EMG signal preprocessing framework tailored for EMG-driven MS modeling of upper-limb motion. The pipeline incorporated muscle-specific spectral filtering based on FFT/PSD analysis, adaptive envelope extraction, and a dual-stage normalization strategy combining maximum voluntary contraction (MVC) and functional muscle weighting.

When applied to elbow flexion–extension simulations using OpenSim, the proposed method achieved high agreement with reference Xsens kinematic data. The most accurate simulation resulted in an average RMSE of 7.78°, MAE of 6.72°, correlation coefficient of 0.98, and minimal phase shift (+2.6%). These results outperformed traditional approaches based solely on static MVC normalization or fixed filter parameters.

This adaptive EMG preprocessing framework enables generation of physiologically accurate and movement-specific excitation profiles suitable for forward dynamics simulation. The method is especially relevant for personalized rehabilitation technologies, EMG-based exoskeleton or prosthesis control, and clinical biomechanical assessment, where individual-specific modeling accuracy is critical.

Footnotes

Ethical considerations

No ethical approval was required.

Consent to participate

Consent to participate was verbal.

Consent for publication

Not applicable.

Author contributions

Conceptualisation, D.C., J.Ž. and K.D.; methodology, D.C. and J.Ž.; software, D.C.; validation, D.C.; data curation, D.C.; writing-original draft preparation, D.C. and J.Ž.; writing-review and editing, J.Ž. and K.D.; visualization, D.C.; supervision, K.D.; project administration, D.C. All authors have read and agreed to the published version of the manuscript.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request.

References

Vega

Shourijeh

, et al. EMG-driven musculoskeletal model calibration with estimation of unmeasured muscle excitations via synergy extrapolation. Front Bioeng Biotechnol 2022; 10: 1–23.

Trigili

Grazi

Crea

, et al. Detection of movement onset using EMG signals for upper-limb exoskeletons in reaching tasks. J Neuroeng Rehabil 2019; 16: 1–16.

Samarakoon

SMUS

Herath

HMKKMB

Yasakethu

SLP

, et al. Long short-term memory-enabled electromyography-controlled adaptive wearable robotic exoskeleton for upper arm rehabilitation. Biomimetics Epub ahead of print 2025; 10: 06.

Cisnal

Gordaliza

Pérez Turiel

, et al. Interaction with a hand rehabilitation exoskeleton in EMG-driven bilateral therapy: influence of visual biofeedback on the Users’ performance. Sensors Epub ahead of print 2023; 23: 2048.

Tahmid

Font-Llagunes

Yang

. Upper extremity muscle activation pattern prediction through synergy extrapolation and electromyography-driven modeling. J Biomech Eng 2024; 146: 1–10.

Cho

Barradas

Schweighofer

, et al. Design of an isometric End-point force control task for electromyography normalization and muscle synergy extraction from the upper limb without maximum voluntary contraction. Front Hum Neurosci 2022; 16: 1–14.

Wang

Pan

. An enhanced EMG-driven musculoskeletal model based on non-negative matrix factorization. Biomed Signal Process Control 2023; 79: 104178.

Barjavel

Pierella

. A realistic upper-limb musculoskeletal model for overall muscle activation estimation. J Biomech 2023; 146: 0–15.

Abdullah

Hulleck

Katmah

, et al. Multibody dynamics-based musculoskeletal modeling for gait analysis: a systematic review. J Neuroeng Rehabil 2024; 21: 78.

10.

Luca

C De

. The use of surface electromyography in biomechanics. J Appl Biomech 1997; 13: 135–163.

11.

Carvalho

Fernández

del-Ama

, et al. Review of electromyography onset detection methods for real-time control of robotic exoskeletons. J Neuroeng Rehabil 2023; 20: 1–16.

12.

Dong

Zhang

. Robust truncated statistics constant false alarm rate detection of UAVs based on neural networks. Drones 2024; 8: 1–21.

13.

Burden

Trew

Baltzopoulos

. Normalisation of gait EMGs: a re-examination. J Electromyogr Kinesiol 2003; 13: 519–532.

14.

Sartori

Farina

Lloyd

. Hybrid neuromusculoskeletal modeling to best track joint moments using a balance between muscle excitations derived from electromyograms and optimization. J Biomech 2014; 47: 3613–3621.

15.

Zhang

Liu

, et al. Estimation of continuous elbow joint movement based on human physiological structure. Biomed Eng Online 2019; 18: 1–15.

16.

Sarshari

Mancuso

Terrier

, et al. Muscle co-contraction in an upper limb musculoskeletal model: EMG-assisted vs. Standard load-sharing. Comput Methods Biomech Biomed Engin 2021; 24: 137–150.

17.

Tahmid

Font-Llagunes

Yang

. Upper extremity joint torque estimation through an electromyography-driven model. J Comput Inf Sci Eng 2023; 23: 1–9. DOI: 10.1115/1.4056255

18.

Chalard

Belle

Montané

, et al. Impact of the EMG normalization method on muscle activation and the antagonist-agonist co-contraction index during active elbow extension: practical implications for post-stroke subjects. J Electromyogr Kinesiol 2020; 51: 102403.

19.

Nam

Zhang

Chow

, et al. Home-based self-help telerehabilitation of the upper limb assisted by an electromyography-driven wrist/hand exoneuromusculoskeleton after stroke. J Neuroeng Rehabil 2021; 18: 1–18.

20.

Farina

Merletti

Enoka

. The extraction of neural strategies from the surface EMG. J Appl Physiol 2004; 96: 1486–1495.

21.

Rook

JWA

Sartori

Refai

. Towards wearable electromyography for personalized musculoskeletal trunk models using an inverse synergy-based approach. IEEE Trans Med Robot Bionics 2025; 7: 13–19. Epub ahead of print 2024.

22.

Pizzolato

Lloyd

Sartori

, et al. CEINMS: a toolbox to investigate the influence of different neural control solutions on the prediction of muscle excitation and joint moments during dynamic motor tasks. J Biomech 2015; 48: 3929–3936.

23.

Laksono

Kitamura

Muguro

, et al. Minimum mapping from EMG signals at human elbow and shoulder movements into two DoF upper-limb robot with machine learning. Machines 2021; 9: 1–13.

24.

Tanaka

Nambu

Maruyama

, et al. Sliding-Window normalization to improve the performance of machine-learning models for real-time motion prediction using electromyography. Sensors 2022; 22: 5005. Epub ahead of print 2022.

25.

Lloyd

Besier

. An EMG-driven musculoskeletal model to estimate muscle forces and knee joint moments in vivo. J Biomech 2003; 36: 765–776.

26.

Mathieu

Crémoux

Gasq

, et al. A feasibility study of leveraging intermuscular coherence in EMG-driven neuromusculoskeletal modeling to improve muscle moment estimation. Biomed Signal Process Control 2025; 109: 1–12.

27.

Cerfoglio

Lopomo

Capodaglio

, et al. Assessment of an IMU-Based Experimental Set-Up for Upper Limb Motion in Obese Subjects. Sensors 2023; 23: 1–16. DOI: 10.3390/s23229264

28.

Ouyang

Cai

Liu

, et al. An improved wavelet threshold denoising approach for surface electromyography signal. EURASIP J Adv Signal Process 2023; 2023; 66: 1–25. DOI: 10.1186/s13634-023-01066-3

29.

Vega

, et al. A computational method for estimating trunk muscle activations during gait using lower extremity muscle synergies. Front Bioeng Biotechnol 2022; 10: 1–18.

30.

Sartori

Reggiani

Farina

, et al. EMG-Driven Forward-Dynamic estimation of muscle force and joint moment about multiple degrees of freedom in the human lower extremity. PLoS One 2012; 7: e52618. Epub ahead of print 2012.

31.

Antonelli

CBB

Hartz

Silva

, et al. The advantages of normalizing EMG to ballistic rather than isometric or isokinetic tasks. J Appl Biomech 2017; 33: 189–196. Epub ahead of print 2016.

32.

Holzbaur

KRS

Murray

Delp

. A model of the upper extremity for simulating musculoskeletal surgery and analyzing neuromuscular control. Ann Biomed Eng 2005; 33: 829–840.

33.

Manal

Gravare-Silbernagel

Buchanan

. A real-time EMG-driven musculoskeletal model of the ankle. Multibody Syst Dyn 2012; 28: 169–180.

34.

Merletti

Farina

. Physiology of muscle activation and force generation. Surf Electromyogr Physiol Eng Appl 2016; 26: 1–29.

35.

Hermens

Freriks

Catherine Disselhorst-Klug

. Development of recommendations for SEMG sensors and sensor placement procedures. J Electromyogr Kinesiol 2000; 810: 361–374.

36.

Leedham

Dowling

. Force-length, torque-angle and EMG-joint angle relationships of the human in vivo biceps brachii. Eur J Appl Physiol Occup Physiol 1995; 70: 421–426.

37.

Seth

Hicks

Uchida

, et al. Opensim: simulating musculoskeletal dynamics and neuromuscular control to study human and animal movement. PLoS Comput Bio 2018; 1: 152–154.