A hybrid EMG–EEG interface for robust intention detection and fatigue-adaptive control of an elbow rehabilitation robot

Abstract

Accurate detection of user intention is a critical requirement for intelligent control systems in upper-limb rehabilitation robots. However, electromyography (EMG)-based recognition can degrade significantly under muscle fatigue. To address this limitation, we propose a hybrid EMG–electroencephalography (EEG) control framework that adaptively fuses peripheral (EMG) and central (EEG) biosignals for robust classification of elbow flexion and extension tasks. The system integrates a support vector machine (SVM)-based EMG classifier and a Common Spatial Pattern (CSP)–SVM EEG classifier, combined through a Bayesian fusion strategy whose weights are modulated in real time according to fatigue levels estimated from EMG spectral features via a k-nearest neighbors (k-NN) model. The hybrid framework was deployed on a lightweight robotic rehabilitation platform and evaluated with five healthy participants (3 females, age 26–39). Results show that adaptive fusion significantly outperformed unimodal baselines, achieving 94.5% classification accuracy (vs. 88.5% for EMG-only) with an end-to-end latency below 500 ms. Importantly, the fatigue-aware weighting preserved performance during high-fatigue conditions (91.4% vs. 83.1% for EMG-only), enhancing system robustness during prolonged sessions. These findings demonstrate the feasibility of a scalable, real-time, fatigue-adaptive control strategy with strong potential for clinical stroke rehabilitation and motor recovery.

Introduction

Upper-limb motor impairments, often resulting from conditions such as stroke, significantly impact individuals’ ability to perform daily activities, thereby affecting their quality of life. Rehabilitation robotics has emerged as a promising solution to facilitate motor recovery by providing consistent, repetitive, and task-specific training^1–4. A key determinant of the efficacy of these robotic systems is the accurate detection of user intent, enabling responsive and adaptive assistance during rehabilitation exercises^5,6.

Unlike previous hybrid EEG–EMG approaches, our work uniquely integrates a continuous, real-time fatigue estimator into a Bayesian fusion framework for intention recognition in rehabilitation robotics. This allows dynamic adaptation to the user’s physiological state, a key limitation of existing systems that assume constant signal quality.

Traditionally, electromyography (EMG) has served as a primary modality for capturing peripheral neuromuscular activity, providing critical insights into voluntary muscle contractions and motor execution. Its non-invasiveness, real-time responsiveness, and compatibility with wearable systems make it particularly suitable for assistive technologies and robotic rehabilitation systems. However, EMG-based control suffers from several well-documented limitations. Most notably, EMG signals are highly susceptible to physiological variability, muscle fatigue, skin impedance, and electrode displacement, which can significantly degrade classification performance and system stability over time^7–9. Furthermore, EMG is inherently peripheral in nature, providing no insight into motor intention or planning before actual movement occurs, limiting their anticipatory control capabilities in rehabilitation contexts¹⁰. In contrast, electroencephalography (EEG) captures cortical brain activity and offers a complementary approach by detecting motor intention through patterns such as motor imagery (MI) or movement-related cortical potentials (MRCPs). EEG enables early-stage decoding of voluntary movement by monitoring activity in the sensorimotor cortex, making it a valuable modality for pre-movement intention detection^11,12. Despite its strengths, EEG-based control also presents challenges. EEG signals are often low in signal-to-noise ratio (SNR), prone to motion artifacts, and affected by user fatigue, mental workload, and individual variability^13,14. Additionally, EEG systems typically require more extensive signal processing and longer training time for effective decoding, which may limit their practicality in fast-paced or precision-critical rehabilitation scenarios^15,16.

To overcome the limitations of single-modality systems, hybrid EEG–EMG control paradigms have emerged as a powerful alternative. By combining EEG’s capacity to detect high-level motor intent with EMG’s ability to reflect fine-grained muscular activity, hybrid systems achieve more robust, accurate, and adaptive user-intent detection^17–20. Such integration enhances system resilience to signal artifacts and enables dynamic compensation during variable movement or fatigue conditions. Recent studies demonstrate that hybrid approaches significantly outperform unimodal systems in classification accuracy, response latency, and user adaptability in real-world and clinical rehabilitation scenarios^21,22. These multimodal systems are particularly advantageous in fatigue-aware control, where EEG can compensate for the degradation of EMG signals over prolonged use, making them well-suited for next-generation human–machine interfaces.

Hybrid EEG–EMG control paradigms, which combine cortical planning (EEG) with peripheral execution signals (EMG), enhance classification robustness, compensate for artifacts, and enable early intention detection. Prior works have demonstrated that combining EEG (reflecting cortical motor planning) with EMG (capturing muscle activation) enhances classification robustness, compensates for signal artifacts, and enables early intention detection. For example, He et al. proposed the EEG-EMG FAConformer architecture using attention-based fusion for robust multimodal pattern recognition²³, while Wang et al. designed a hybrid Brain-Machine-Interface (BMI) system that dynamically switches between modalities to mitigate fatigue-induced errors²⁴. While these studies demonstrate progress in multimodal fusion, they either rely on static integration or modality switching, or do not explicitly adapt to fatigue-induced signal variability in real time. Our approach differs by embedding a fatigue-aware Bayesian weighting mechanism directly into the fusion process, ensuring robust decoding under prolonged use. Other studies have used deep learning frameworks^21,22 to improve classification accuracy and response stability. However, these systems often lack mechanisms to adapt to fatigue-induced variability in real-time, particularly in rehabilitation contexts. Our proposed system addresses this gap by integrating real-time fatigue estimation into the fusion process, offering a scalable and physiologically responsive solution for upper-limb motor intention detection.

In rehabilitation contexts, muscle fatigue is not simply a source of noise but a critical physiological factor that directly impacts both user safety and system performance. Prolonged sessions can degrade EMG signal quality and alter cortical activity patterns, leading to misclassification of user intention and reduced therapeutic effectiveness. Without mechanisms for real-time adaptation, rehabilitation robots risk delivering inappropriate levels of assistance or causing user discomfort, thereby limiting long-term usability. For these reasons, incorporating fatigue-awareness into the control framework is essential to ensure resilience, safety, and sustained performance in dynamic operational environments. This need for physiologically responsive adaptation directly motivates the fatigue-aware hybrid control strategy proposed in this work.

Despite these advancements, significant challenges remain in the seamless integration of EMG and EEG signals, particularly when it comes to dynamically balancing their contributions under real-time physiological variability such as muscle fatigue. Fatigue alters both peripheral muscle signals and cortical activity patterns, leading to substantial degradation in unimodal performance. Existing hybrid systems typically rely on static fusion strategies, which are unable to adapt to these fluctuations, thereby limiting robustness and long-term usability. Our work directly addresses this gap by introducing an adaptive fusion framework in which the relative weighting of EMG and EEG inputs is continuously modulated based on real-time fatigue estimation. This physiologically responsive mechanism enables more reliable intention detection and ensures consistent performance during prolonged rehabilitation sessions, representing a key contribution beyond prior hybrid approaches.

To the best of our knowledge, no previous work has integrated a real-time fatigue estimator into a hybrid EMG–EEG control framework for rehabilitation robotics. While earlier studies have explored hybrid EMG–EEG intention recognition^25,26, these approaches relied on static or predefined fusion schemes and did not consider physiological variability induced by fatigue. Likewise, Bayesian fusion strategies have been applied in other contexts, but they have not been combined with a dynamic, fatigue-aware modulation of modality weights. The novelty of this study therefore lies in introducing a physiologically responsive Bayesian fusion mechanism that continuously adapts the contributions of EMG and EEG signals according to the user’s fatigue state, ensuring robust and safe intention detection during prolonged rehabilitation sessions.

The main contributions are:

• Development of a hybrid EMG–EEG control system for a portable elbow rehabilitation robot.

• Design of a k-NN–based fatigue estimator operating on EMG spectral features in real time.

• Integration of the fatigue estimator into an adaptive Bayesian fusion strategy that dynamically adjusts EEG–EMG weighting.

• Experimental validation showing improved robustness, safety, and usability compared to static fusion methods.

The structure of this paper is organized as follows: section “Materials and methods” details the materials and methodologies used in this study, section “Results and discussion” presents the findings and corresponding discussions, and the “Conclusions” provides a summary and perspective on the contributions and future directions of this work.

Materials and methods

Approach

In this work, we propose a novel adaptive hybrid EMG–EEG control system designed to detect user intention during upper-limb rehabilitation, with integrated real-time muscle fatigue estimation. Our method leverages the complementary strengths of surface electromyography (sEMG) and electroencephalography (EEG) signals to robustly decode motor intentions (elbow flexion and extension), while dynamically adjusting to neuromuscular fatigue.

The proposed system consists of the following modules (Fig. 1):

Fig. 1.

Block diagram of the proposed hybrid EMG–EEG control framework for elbow rehabilitation. The system includes signal acquisition (EEG and EMG), parallel intention detection modules (CSP + SVM and EMG + SVM), where PEMG(y) = probability of class label y from the EMG classifier and PEEG(y) = probability of class label y from the EEG classifier, real-time fatigue estimation via k-NN, and adaptive Bayesian fusion that modulates final decision output.

• Signal Acquisition.

  • EMG signals are acquired using a wearable armband from elbow-related muscles (biceps brachii, triceps brachii).
  • EEG signals are captured using OpenBCI from motor cortex electrodes.

• Intention Classification.

  • EMG signals are preprocessed and classified via a Support Vector Machine (SVM) model to predict user intention.
  • EEG signals are filtered, spatially enhanced via Common Spatial Pattern (CSP), and classified using another SVM model.

• Muscle Fatigue Estimation.

  • EMG features are input to a modified k-NN classifier that computes a continuous fatigue score f(x)∈[0,1].

• Bayesian Fusion Module.

  • The EMG and EEG classifier probabilities are combined via an adaptive Bayesian rule.
  • The fusion weight adjusts the influence of EEG based on fatigue level.

• Decision Output.

  • The fused intention (flexion or extension) is used as the final control signal for actuation (e.g., robotic arm movement or rehabilitation task progression).

In this work, we adopted a decision-level fusion strategy for combining EEG and EMG signals. This choice was motivated by the heterogeneous nature of the two modalities: EEG requires spatial filtering (e.g., CSP) and operates at lower SNR, while EMG offers faster response and localized activity. By classifying each modality independently and then integrating the predicted probabilities, the system maintains modular preprocessing pipelines and minimizes inter-modality interference. Additionally, decision-level fusion allows the fusion weights to be dynamically adjusted based on real-time muscle fatigue, which would be more difficult to achieve in earlier fusion levels. This modular and adaptive design improves flexibility, robustness, and physiological relevance, particularly in rehabilitation scenarios where neuromuscular states vary over time.

Rehabilitation robot

The proposed hybrid EMG–EEG control system was integrated with a portable, lightweight robotic rehabilitation device designed specifically for upper-limb rehabilitation, with a focus on elbow joint movements (flexion and extension). This integration enables real-time intention-driven control, ensuring adaptability to patient fatigue while delivering precise robotic assistance. Figure 2 presents the overall system architecture, including the rehabilitation robot structure and the EMG electrode placement used during the experiments.

Fig. 2.

Schematic architecture of the elbow rehabilitation robot and EMG electrode placement over the biceps and triceps muscles. This figure illustrates the experimental configuration but does not include participant photographs, in order to respect ethical and privacy considerations.

The robotic system is adapted from our previous work^1,27, and incorporates a detachable, bilateral arm support that accommodates both right and left arms. A central goal of the design is to achieve precise anatomical alignment with the user’s elbow joint, ensuring biomechanically correct motion guidance during rehabilitation sessions.

To support anatomical variability across users, the system employs adjustable outer plates for both the upper arm and forearm. These plates form the core structural framework and include strategically distributed holes to facilitate the attachment of stepper motors and rotary bearings. A total of 12 precision bearings per plate (mounted 12 mm above and below the inner plate) enable smooth and friction-free articulation of the elbow joint.

Customization is achieved through a screw-nut translation mechanism, driven by two stepper motors, which adjusts the linear positioning of the arm and forearm plates. This mechanism ensures the exoskeleton can adapt to a wide range of user arm lengths and anthropometric profiles, while maintaining the rigid alignment necessary for effective rehabilitation and consistent signal interpretation from EMG/EEG sources.

The robotic system is passively supported during calibration and actively driven during execution of elbow flexion and extension movements. Control commands are triggered by the fused output of the EMG and EEG classifiers, with the system adapting to fatigue levels in real-time.

Biosignals acquisition

EEG acquisition system

In this study, electroencephalographic (EEG) signals were acquired using the OpenBCI Cyton board, an open-source, modular, and wireless bio-signal acquisition platform capable of recording high-resolution EEG data²⁸. The Cyton board provides 8 channels with a sampling rate of 250 Hz, which is sufficient for capturing motor-related EEG dynamics, particularly those associated with upper-limb movement intentions.

EEG electrodes were placed following the international 10–20 system, targeting regions over the sensorimotor cortex known to exhibit event-related desynchronization/synchronization (ERD/ERS) during voluntary or imagined movement. Specifically, electrodes were placed at C3, Cz, and C4 to monitor bilateral motor activity corresponding to the left, midline, and right sensorimotor cortices, respectively. To improve spatial resolution and facilitate CSP filtering, additional electrodes were placed at FC3, FC4, CP3, and CP4, covering the frontal and parietal regions adjacent to the central motor strip.

The EEG reference electrode was positioned at the left earlobe (A1) to minimize contamination from motor cortical activity, while the ground (bias) electrode was placed at Fpz, the midline frontal scalp position. This common reference montage setup supports effective spatial filtering and classification performance when combined with CSP and SVM-based algorithms.

Prior to each recording session, electrode-skin impedance was reduced to below 10 kΩ using conductive gel to ensure high signal quality. The EEG data were bandpass filtered within the 8–30 Hz range to isolate the mu (8–13 Hz) and beta (13–30 Hz) rhythms, which are strongly modulated during upper-limb motor execution and intention tasks.

This configuration provides reliable and repeatable acquisition of cortical signals related to elbow flexion and extension, forming the EEG input branch of the proposed hybrid control system.

Figure 3 illustrates the electrode placements over the motor cortex according to the 10–20 system:

Fig. 3.

Electrode locations on the 10–20 system.

EMG acquisition system

sEMG signals were acquired using a wearable EMG armband, which offers a practical and non-invasive approach for capturing muscle activity associated with upper-limb movement. The device is equipped with 8 evenly spaced dry electrodes placed circumferentially around the forearm, enabling simultaneous recording from multiple forearm muscles involved in elbow flexion and extension, including the biceps brachii, brachioradialis, and triceps brachii.

The armband samples sEMG signals at 200 Hz and includes onboard signal conditioning such as amplification, rectification, and initial filtering²⁹. For this study, raw EMG data were further processed offline using a 4th-order Butterworth band-pass filter (20–150 Hz) to eliminate low-frequency motion artifacts and high-frequency noise, preserving the core frequency components of voluntary muscular activity³⁰.

The sensor placement is designed for ease of use and repeatability without requiring precise anatomical targeting. However, alignment was maintained across sessions to ensure consistent signal capture from key muscle groups responsible for the intended joint motion.

Participants performed repeated trials of elbow flexion and extension, during which EMG signals were recorded and labeled. These signals served as the input to a SVM classifier trained to distinguish between the two motor intentions based on muscle activation patterns. Features were extracted from segmented EMG windows to construct the input vector for classification.

This setup provides a reliable and lightweight means of detecting user motor intention from peripheral muscle activity and forms the EMG branch of the hybrid classification system.

EEG feature selection for intention detection (SVM + CSP)

To decode user motor intention (elbow flexion vs. extension) from EEG signals, a combination of CSP and SVM classification was employed. This approach has been widely validated for motor imagery and motor execution tasks, as it efficiently extracts discriminative features from multi-channel EEG signals recorded over the motor cortex³¹.

Common Spatial patterns (CSP)

The CSP algorithm is a supervised spatial filtering technique that enhances class-specific variance while suppressing variance common to both classes. It projects multichannel EEG data onto spatial filters that maximize the difference in signal variance between two conditions—in this case, elbow flexion and extension. Mathematically, CSP solves a generalized eigenvalue problem to obtain projection matrices that separate the two classes. The EEG data matrix , where C is the number of channels and T is the number of time samples, is transformed using the CSP filter W to obtain spatially filtered signals .

From these transformed signals, the log-variance of the most discriminative CSP components is computed to construct the feature vector

1

where is the i^th CSP feature. In this study, four CSP components (two from each side of the spectrum) were retained, yielding a compact and robust representation of motor intention.

Support vector machine classification

The extracted CSP features were classified using an SVM with a linear kernel, which has been shown to perform effectively for EEG-based binary classification problems³². The linear SVM finds a hyperplane that maximizes the margin between classes in the feature space. Given the low dimensionality and near-linear separability of CSP features, a linear kernel provides an optimal trade-off between classification accuracy and computational efficiency.

The SVM was trained using a 5-fold cross-validation scheme to avoid overfitting and ensure generalizability. Accuracy, sensitivity, and specificity metrics were used to evaluate performance across trials. All signal preprocessing and classification were implemented in Python using the MNE, scikit-learn, and OpenBCI libraries.

This EEG feature selection pipeline (Fig. 4) provides a robust input to the Bayesian fusion framework by producing reliable and real-time estimates of motor intention from cortical activity.

Fig. 4.

EEG feature selection pipeline.

EMG feature selection for intention detection (SVM)

To detect the user’s motor intention—specifically elbow flexion and extension—sEMG signals were extracted from forearm muscles most directly involved in this movement. Using a multi-channel EMG armband, we selectively processed signals from electrodes consistently aligned with the biceps brachii and triceps brachii, which are the primary agonist-antagonist pair for elbow motion.

Targeted muscle signal selection

Among the armband’s 8 sensors, two channels were consistently identified to overlay the biceps region (responsible for elbow flexion) and the triceps region (responsible for elbow extension). Figure 5 shows the approximate electrode placement on the forearm, indicating the selected channels targeting the biceps brachii (flexor) and triceps brachii (extensor) muscles. These channels were selected based on anatomical alignment and confirmed through pilot muscle activation tests during flexion and extension exercises. Channels with low or noisy activation were excluded from further processing.

Fig. 5.

Electrode placement of selected EMG channels targeting the biceps and triceps muscles for elbow flexion and extension.

This selective approach reduces feature dimensionality, minimizes redundancy, and emphasizes physiologically meaningful information—improving classification interpretability and efficiency.

Preprocessing and Wndowing

The selected EMG signals were band-pass filtered using a 4th-order Butterworth filter (20–150 Hz) to isolate the relevant muscle activation frequency range and suppress noise. Filtered data were segmented into 250 ms windows with 50% overlap, capturing dynamic muscular events with low latency.

Feature extraction

Each window was transformed into a low-dimensional feature vector using established time-domain features known for their discriminative capacity in EMG-based motion classification^33,34:

• Root Mean Square (RMS): Measures signal power and muscle contraction intensity.

• Waveform Length (WL): Captures signal complexity and temporal variation.

• Zero Crossing Rate (ZCR): Reflects frequency characteristics of motor activity.

• Slope Sign Changes (SSC): Sensitive to rapid changes in activation profiles.

Features were computed individually for each selected channel, and then concatenated to form a compact input vector for classification.

SVM-based intention classification

An SVM with a Radial Basis Function (RBF) kernel was employed to classify each feature vector into one of two categories: flexion or extension. This classifier was chosen for its robustness in modeling nonlinear relationships in EMG features. Hyperparameters (C and γ) were optimized via grid search with 5-fold cross-validation on the training data.

This classifier provides a high-confidence, low-latency prediction of user intention from peripheral muscle activity. Its output feeds directly into the Bayesian fusion framework, where it is fused with a parallel EEG-based prediction to produce the final decision.

Figure 6 illustrates a summary of EMG intention detection pipeline.

Fig. 6.

EMG intention detection pipeline.

Muscle fatigue estimation via k-NN on EMG signals

To enable dynamic adaptation of the fusion process in response to physiological changes, we implemented a real-time muscle fatigue estimation system based solely on sEMG signals (Fig. 7). The goal is to detect the onset and progression of localized muscular fatigue during repetitive elbow flexion and extension tasks and use this information to adjust the confidence weighting of the EMG classifier within the Bayesian fusion model.

Fig. 7.

Fatigue estimation pipeline.

Fatigue-related feature extraction

Muscle fatigue induces characteristic changes in EMG signal properties, including reductions in mean and median frequency, and increases in amplitude and spectral compression^33,35. We extracted the following five features: Mean Frequency (MNF), Median Frequency (MDF), Root Mean Square (RMS), Zero Crossing Rate (ZCR), and the slope of MDF/MNF over time. These features were computed per channel and averaged across biceps and triceps to form a compact fatigue feature vector. To quantify these effects, the following time-domain and frequency-domain features were extracted from each EMG window:

• MNF: Downward shift indicates fatigue.

• MDF: Sensitive to conduction velocity changes.

• RMS: Increases as motor units compensate for fatigue.

• ZCR: Reflects spectral compression.

• Slope of MDF/MNF over time: Trend-based fatigue indicator.

These features were computed per channel and averaged across biceps and triceps signals to obtain a compact feature vector reflecting global muscle state.

k-NN-based fatigue classification

The extracted features were used to train a k-NN classifier. The model assigns a fatigue level to each time window based on proximity to labeled examples in the feature space. For this study, fatigue was classified into three levels:

1. Low (Non-Fatigued).

2. Moderate.

3. High (Fatigued).

The model was trained using annotated EMG datasets collected from subjects performing prolonged flexion-extension cycles until voluntary fatigue. Euclidean distance was used as the similarity metric, and k = 5 was selected via cross-validation to balance sensitivity and robustness.

Integration with adaptive control

The real-time fatigue level output from the k-NN model is used to modulate the fusion weights in the Bayesian decision layer. As muscle fatigue increases, the EMG classifier’s confidence is reduced relative to the EEG stream, thereby maintaining decision stability and improving resilience to physiological degradation.

Adaptive bayesian fusion with continuous fatigue-aware weighting

To achieve robust and context-sensitive multimodal intention recognition, we employ an adaptive Bayesian fusion framework that dynamically adjusts the contribution of EMG and EEG classifiers based on the user’s real-time muscle fatigue level. This section details the use of a continuous fatigue score to regulate the fusion process.

Fusion model overview

The system combines classifier outputs from two modalities:

: Probability of class label y∈ {Flexion, Extension} from the EMG-based SVM.

: Probability from the EEG-based CSP + SVM classifier.

The final classification result is computed using a weighted sum:

2

Where is the adaptive fusion coefficient modulated by the user’s current fatigue level.

Continuous fatigue score from k-NN

To enable fine-grained adaptation, we modified the standard k-NN classifier to estimate fatigue as a continuous score Fatigue labels from training data are encoded numerically:

• Low Fatigue: f(x) < 0.33.

• Moderate Fatigue: 0.33 ≤ f(x) < 0.66.

• High Fatigue: f(x) ≥ 0.66.

The thresholds were determined empirically based on the distribution of subjective fatigue ratings and EMG spectral shifts collected during pilot trials. Specifically, we mapped continuous fatigue scores to three categories (low, moderate, high) by aligning them with Borg scale ratings (1–5) reported by participants. The cut-off values (< 0.33, 0.33–0.66, ≥ 0.66) correspond to tertile boundaries of normalized fatigue scores.

Given a new feature vector x, the fatigue score is computed as a distance-weighted average over the k nearest neighbors:

3

is the fatigue level of neighbor i

​ is the Euclidean distance to neighbor i

is a small constant to avoid division by zero.

This produces a smooth fatigue score that evolves continuously as the user experiences fatigue.

Mapping fatigue to fusion weight

The EEG fusion weight α is then computed as a linear function of fatigue:

4

The two hyperparameters in Eq. (4) were determined through an offline grid search over the training dataset. Candidate values were evaluated based on their impact on the overall classification accuracy across different fatigue levels. The final values were selected to balance sensitivity to fatigue variation with stability in fusion performance. This empirical calibration ensures the adaptive function accurately reflects changes in muscle fatigue while maintaining robustness in intention recognition.

This function ensures:

• When fatigue is low → prioritizes EMG.

• When fatigue is high → prioritizes EEG.

Such adaptive weighting enhances the robustness of intention detection under varying neuromuscular conditions.

The final decision is made as:

5

This fusion framework provides a real-time adaptive interface that balances between muscle-driven and brain-driven control, compensating for performance degradation due to fatigue.

Results and discussion

Experimental setup

This section describes the experimental protocol used to evaluate the proposed hybrid EMG–EEG control system, including participant recruitment, signal acquisition, task design, and integration with a robotic rehabilitation device targeting elbow joint rehabilitation.

Participants

Five neurologically healthy adult participants (3 females, 2 males; age range: 26–39 years) were recruited for this study. All participants were right-handed, with no history of musculoskeletal or neurological impairments. Informed consent was obtained from all participants in accordance with ethical standards. The experimental protocol was conducted in full compliance with international ethical standards, including the Declaration of Helsinki, and was formally reviewed and approved by the Scientific Research Ethics Committee of the Taif University (Approval No. 47 − 046), which is accredited by the National Committee for Bioethics with No. (HAPO-02-T-105).

Task design

Participants were seated with their dominant arm fitted into the robotic system. They were instructed to perform alternating elbow flexion and extension movements in response to visual cues displayed on a screen. The robotic system moved in synchronization with decoded user intent.

Each subject performed a total of 40 trials, consisting of 20 elbow flexion and 20 elbow extension tasks. These were divided across two recording sessions with adequate rest between blocks. In total, 200 trials were collected from five participants. The dataset was used for feature extraction, classification training, and validation of the fatigue-adaptive fusion framework.

For each trial, both EEG and sEMG signals were segmented into 3-second windows starting from the onset of the visual cue. Temporal alignment between modalities was ensured through synchronized data acquisition, with both EEG and EMG systems triggered simultaneously via the LabVIEW interface. All samples were timestamped using a shared system clock to eliminate misalignment. This approach ensured that the EEG and EMG signals corresponded to the same motor imagery or contraction task window for reliable fusion analysis. For feature extraction, the EMG data within each 3-second trial were further processed using sliding 1-second windows with 50% overlap, as shown in Fig. 7. EEG features were extracted from the full 3-second window using the CSP + SVM pipeline.

Prior to the experiment, each participant completed a short training session to learn and practice the motor imagery (MI) of elbow flexion and extension. Standardized verbal instructions were provided, emphasizing kinesthetic imagery (i.e., imagining the sensation of the movement rather than visualizing it). During the actual trials, visual cues indicated which movement to imagine and when to start. Participants were instructed to remain completely still during MI trials. A trained experimenter supervised the sessions to ensure compliance, attention, and consistency, and issued verbal reminders or pauses if movement artifacts or lapses in focus were observed.

Ground truth and labeling

Ground-truth labels for movement intention were defined based on the visual cue sequence and confirmed using kinematic tracking from the robot’s built-in sensors. For fatigue annotation, a simplified labeling approach was adopted based on session timing and subjective user ratings on a 5-point perceived exertion scale. These ratings were discretized into Low, Moderate, and High fatigue levels for supervised training of the fatigue classifier.

Experimental results and evaluation

This section presents the performance evaluation of the proposed hybrid EMG–EEG control system for elbow flexion and extension detection, incorporating muscle fatigue adaptation via Bayesian fusion. Results are reported for individual modalities, the fusion model, and the impact of fatigue-based adaptation.

To assess classification performance, the following metrics were computed:

• Accuracy (Acc): Overall correct predictions over total predictions.

• Precision (Pre) and Recall (Rec): Per-class performance on Flexion and Extension.

• F1-Score: Harmonic mean of precision and recall.

• Confusion Matrix: To visualize classification tendencies.

• Response Latency: Time from signal acquisition to classification output.

• Fatigue Mapping Consistency: Smoothness and temporal reliability of fatigue score.

Standard deviations (± SD) are reported for all accuracy values in Table 1. These were computed over 5-fold cross-validation across all five participants to represent inter-subject and intra-subject variability.

Table 1.

Comparative classification accuracy (%) of intention detection under different fatigue conditions.

Condition	EMG Only	EEG Only	Static fusion	Adaptive fusion
Low fatigue	92.6 ± 3.1%	86.3 ± 4.2%	94.1 ± 2.6%	94.5 ± 2.0%
High fatigue	83.1 ± 5.2%	85.6 ± 4.8%	89.2 ± 3.9%	91.4 ± 2.8%
Avg.	88.5 ± 4.2%	85.9 ± 4.5%	91.7 ± 3.3%	93.0 ± 2.4%

Fatigue levels were computed from EMG signals using a continuous fatigue score, discretized into low (< 0.33), moderate (0.33–0.66), and high (≥ 0.66) categories.

The EMG-based SVM classifier, utilizing signals from the biceps and triceps muscles, demonstrated high performance in distinguishing between elbow flexion and extension. It achieved an average accuracy of 92.6% with a standard deviation of ± 3.1%, indicating consistent classification across participants. The F1-score was 0.93 for flexion and 0.91 for extension, reflecting balanced precision and recall for both movement classes. Additionally, the system exhibited a low latency of approximately 180 milliseconds per prediction window, ensuring near real-time responsiveness. These results highlight the strong discriminative capability of localized EMG signals for detecting user intention in upper-limb motor tasks.

The EEG-based intention detection pipeline, employing CSP for feature extraction followed by classification using an SVM, achieved reliable performance in distinguishing between elbow flexion and extension. Across participants, the system reached an average classification accuracy of 86.4% (± 4.7%), with the F1-scores of 0.85 for flexion and 0.87 for extension, indicating strong class-wise discrimination. While slightly lower than the EMG-only classifier, the EEG-based method provided valuable complementary information, particularly in cases where EMG signals were affected by fatigue or noise. The classification latency averaged around 250 milliseconds due to CSP computation overhead, but remained within acceptable limits for real-time operation. These findings confirm that EEG signals, when processed with spatial filtering and machine learning, offer robust support for user intention decoding in rehabilitation contexts.

The k-NN-based muscle fatigue classifier demonstrated effective performance in estimating continuous fatigue levels from EMG signals, using both spectral and amplitude features. When evaluated across three discrete fatigue levels (low, moderate, and high), the classifier achieved an average classification accuracy of 88.2% with a standard deviation of ± 4.9%. Furthermore, the continuous fatigue score output showed a strong correlation with participants’ subjective fatigue ratings (ρ = 0.82), validating the physiological relevance of the estimated fatigue metric. Importantly, the fatigue scores exhibited stable temporal behavior with no abrupt fluctuations, which is critical for ensuring smooth and reliable adaptation of the Bayesian fusion parameters. These results confirm that the proposed EMG-based approach enables robust real-time tracking of muscle fatigue, supporting adaptive modulation of the control system.

The adaptive Bayesian fusion framework effectively combined the classification results from the EMG and EEG pipelines, with dynamic weighting modulated by the real-time fatigue score estimated from EMG features. Under low fatigue conditions, the system increased the weight of the EMG classifier, leveraging its higher precision and lower latency. Conversely, under moderate to high fatigue, the EEG contribution was upweighted to compensate for potential EMG signal degradation. This adaptive weighting strategy led to a significant improvement in intention detection accuracy compared to either modality alone, yielding an overall fused accuracy of 93.0% ±2.4%, as shown by Table 1. The continuous fatigue score f(x)∈[0,1] was discretized into three categories for comparative analysis: Low Fatigue (f(x) < 0.33, Moderate Fatigue (0.33 ≤ f(x) < 0.66), and High Fatigue (f(x) ≥ 0.66). A one-way repeated measures ANOVA revealed statistically significant differences in classification accuracy among the four approaches (p < 0.05). The adaptive Bayesian fusion method outperformed EMG-only, EEG-only, and static fusion strategies. However, due to the small sample size (n = 5), these results should be interpreted with caution. Future studies with larger cohorts are needed to validate these findings. The fusion process remained stable across trials, with no evidence of oscillatory weighting or performance drop-offs during fatigue transitions. These results highlight the effectiveness of incorporating physiological state awareness into the fusion model, enhancing robustness and reliability in real-time motor intention decoding for upper-limb rehabilitation applications. Overall, these findings should be considered preliminary; validation with larger clinical cohorts is required to confirm robustness and generalizability.

Statistical analysis using one-way ANOVA with repeated measures confirmed that the improvements in classification accuracy achieved by the adaptive Bayesian fusion method were statistically significant (p < 0.05) compared to EMG-only, EEG-only, and static fusion methods under both low and high fatigue conditions.

Average system response time (EMG + EEG + fusion) was < 500 ms, making it suitable for real-time control. Participants reported no discomfort, and robotic movement was described as smooth and intention-aligned. Fatigue levels were perceived to be accurately reflected in system behavior.

To validate the behavior of the adaptive fusion weight α(f), we analyzed its variation across different fatigue conditions. Figure 8 shows the average α values computed for low, moderate, and high fatigue levels, based on the normalized fatigue scores using Eq. (4). The results confirm that α decreases progressively with increasing fatigue, indicating a reduced reliance on EMG inputs under fatigue and greater weighting of EEG signals. This trend supports the intended behavior of the Bayesian fusion strategy.

Fig. 8.

Variation of the fusion weight α(f) with muscle fatigue level. Mean ± SD of α values are shown for low, moderate, and high fatigue conditions across all subjects.

Figure 9 presents the confusion matrices corresponding to the classification performance of the proposed system under four different configurations: EMG-only, EEG-only, static fusion, and adaptive Bayesian fusion. These visualizations provide a clear comparison of true and predicted labels for elbow flexion and extension tasks, illustrating the impact of modality and fusion strategy on classification accuracy and robustness. To better interpret classification performance, the confusion matrix in Fig. 9 has been normalized row-wise. This highlights the classifier’s tendency to correctly identify flexion and extension tasks and reveals any confusion patterns. The adaptive Bayesian fusion approach shows notably higher consistency and fewer misclassifications, especially under varying fatigue conditions.

Fig. 9.

Normalized confusion matrices illustrating classification performance for elbow flexion and extension across EMG, EEG, Static Fusion, and Adaptive Bayesian fusion approaches.

Figure 10 illustrates the classification accuracy of user intention (elbow flexion vs. extension) under low and high fatigue conditions using four different approaches: EMG-only, EEG-only, static fusion, and adaptive Bayesian fusion. The bar chart highlights the performance degradation observed in EMG and static fusion methods under fatigue, while the adaptive Bayesian fusion approach maintains superior accuracy by dynamically adjusting the modality contributions based on real-time fatigue estimation. This demonstrates the robustness and adaptability of the proposed fusion strategy in fatigue-sensitive rehabilitation scenarios.

Fig. 10.

Classification accuracy across fatigue levels for EMG, EEG, static fusion, and adaptive Bayesian Fusion Methods.

Figures 11, 12, and 13 present the comparative performance metrics—Precision, Recall, and F1-Score, respectively—for the four classification approaches: EMG-only, EEG-only, static fusion, and adaptive Bayesian fusion. These metrics offer a detailed view of each method’s ability to correctly identify user intentions (elbow flexion or extension). While EMG and EEG individually show competitive scores, the fusion methods—particularly the adaptive Bayesian approach—demonstrate consistent improvements across all metrics. This confirms the effectiveness of the proposed system in enhancing classification reliability, especially under variable muscle fatigue conditions.

Fig. 11.

Precision comparison for EMG, EEG, static fusion, and adaptive Bayesian fusion approaches.

Fig. 12.

Recall comparison for EMG, EEG, static fusion, and adaptive Bayesian fusion approaches.

Fig. 13.

F1-score comparison for EMG, EEG, static fusion, and adaptive Bayesian fusion approaches.

Table 2 summarizes the average classification performance metrics—accuracy, precision, recall, and F1-score—across all tested approaches: EMG-only, EEG-only, static fusion, and adaptive Bayesian fusion. These metrics were computed by averaging results across all participants and both movement classes (elbow flexion and extension), offering a comprehensive comparison of system effectiveness under real-time conditions.

Table 2.

Summary of classification performance metrics (average across classes).

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
EMG only	92.6	92.0	91.5	91.8
EEG only	86.3	86.0	85.0	85.5
Static fusion	94.1	93.5	92.5	93.0
Adaptive fusion	94.5	94.5	94.0	94.2

Overall, the results confirm that the adaptive hybrid EMG–EEG control system, supported by fatigue-aware Bayesian fusion, significantly outperforms unimodal and static fusion methods, demonstrating its suitability for robust and personalized upper-limb rehabilitation.

Discussion

This study introduced and evaluated a hybrid EMG–EEG control system featuring adaptive Bayesian fusion and real-time muscle fatigue estimation, integrated into an upper-limb rehabilitation robot focusing on elbow joint movements. The system’s performance was assessed through experiments involving five neurologically healthy participants, demonstrating its potential in enhancing rehabilitation outcomes.

The novelty of the proposed approach lies in its integration of real-time muscle fatigue estimation into a dynamic decision-level fusion framework for hybrid EMG–EEG control. While prior studies have explored EEG–EMG integration for motor decoding, most rely on static fusion mechanisms and do not account for intra-session physiological variability such as fatigue. By using a fatigue-aware Bayesian model, our system dynamically shifts fusion weights to favor EEG signals under high fatigue conditions. This trial-level adaptability, combined with seamless integration into a real-time robotic control loop, constitutes a novel contribution to the field of neuroadaptive rehabilitation.

Integrating EMG and EEG signals capitalizes on the complementary strengths of both modalities: EMG offers high specificity for localized muscular activation, while EEG provides insight into central motor planning. This complementary integration aligns with findings from²³, who developed the EEG-EMG FAConformer—a multimodal motion pattern recognition algorithm combining EEG and EMG signals using a Frequency Aware Conv-Transformer architecture. Their approach effectively eliminated irrelevant information in EEG and EMG signals, achieving high robustness and stability in motor pattern recognition. our system employs a k-NN-based fatigue classifier and adaptive fusion, emphasizing real-time adaptability and fatigue-aware control, whereas FAConformer focuses on deep learning-based signal fusion for pattern recognition.

The adaptive Bayesian fusion model in our system dynamically adjusts the weighting of EMG and EEG inputs based on real-time fatigue levels, preserving classification accuracy even as EMG signal quality degrades due to muscle fatigue. This adaptive capability is crucial in real-world rehabilitation scenarios where users may experience varying effort and endurance levels throughout sessions. Wang et al. presented a hybrid brain-machine interface integrating EEG and EMG to improve multimodal control and mitigate fatigue in assistive applications²⁴. Their system dynamically alternated between EEG and EMG inputs, achieving task completion times comparable to EMG-only approaches while reducing physical demand. While both systems utilize EEG and EMG integration, our approach focuses on adaptive Bayesian fusion with real-time fatigue estimation, providing continuous control adjustments based on user fatigue levels.

To further validate the contribution of our method, we compared its performance with other recent hybrid EEG–EMG control systems reported in the literature. While deep learning-based architectures²³ and transformer networks offer high classification accuracy, they often lack real-time adaptability to physiological states like muscle fatigue. Similarly, switching-based methods²⁴ offer limited fatigue compensation but lack fine-grained adaptability. In contrast, our method integrates real-time, continuous fatigue estimation with adaptive fusion, achieving higher resilience under dynamic neuromuscular conditions. Our average classification accuracy of 94.5% under variable fatigue conditions is competitive with the state of the art, while maintaining sub-500 ms response latency suitable for real-time control. This comparative performance highlights both the robustness and the novelty of the proposed fatigue-aware fusion framework.

In contrast to systems that rely on discrete fatigue states, our framework employs a k-NN-based classifier for continuous fatigue estimation, enabling fine-grained, real-time control adaptation. This fatigue-aware adjustment enhances both safety and effectiveness in assistive control, reducing the risk of overexertion while improving the relevance of system responses. Notably, Pongsing et al. proposed a novel strategy enabling a rehabilitation robot to generate self-adaptive resistance using EMG-based Fuzzy-PI control, demonstrating improved movement smoothness and extended rehabilitation durations³⁶. Moreover, Wang et al. proposed a fatigue judgment method using multi-information fusion of sEMG, heart rate variability (HRV), and instantaneous heart rate (IHR) with a fuzzy logic controller³⁷. They achieved a fatigue judgment error of 4.3%, satisfying the requirements for fatigue assessment in rehabilitation training. While both systems aim to assess muscle fatigue during rehabilitation. Our system uniquely integrates EEG signals and employs adaptive Bayesian fusion, offering a more comprehensive approach to fatigue-aware control.

The integration of our hybrid system into a lightweight, adjustable rehabilitation robot showed promising results. Participants experienced natural, responsive interaction, and the system reliably recognized user intent in both active and fatigued states. This has strong implications for adaptive rehabilitation training, where long sessions may diminish muscular performance and impair intention detection in traditional EMG-only systems. Sarasola-Sanz et al. validated the feasibility and functionality of a hybrid brain-muscle-machine interface for stroke rehabilitation, demonstrating its potential to potentiate the plasticity of the entire neural system from the brain to the muscles³⁸.

Our hybrid EMG–EEG control system distinguishes itself by integrating real-time fatigue estimation with adaptive Bayesian fusion, providing a responsive and personalized rehabilitation experience. While other studies have explored EEG–EMG integration^39–41 and fatigue assessment^42–44 separately, our approach combines these elements to address the dynamic nature of user fatigue during rehabilitation.

While this study focuses on classical machine learning methods due to their computational efficiency and interpretability, recent literature has shown promising results with deep learning-based hybrid control systems for EEG and EMG decoding^45–47. These methods offer strong modeling capacity but often require large datasets, longer training times, and specialized hardware. In future work, we plan to explore the integration of lightweight deep architectures (e.g., CNNs or transformers) into our adaptive control framework, particularly in combination with the fatigue-aware fusion strategy proposed here.

Despite promising results, one limitation of this study is that the experiments were conducted exclusively on healthy participants. While this was appropriate for initial feasibility validation and algorithm calibration, further testing on clinical populations (e.g., stroke survivors or patients with neuromuscular impairments) is essential to assess the robustness and translational value of the proposed system. A future phase of this work will involve clinical trials under medical supervision. Additionally, while SVM and CSP were selected for EEG processing, deep learning models may offer further improvements and should be explored in subsequent work. Similarly, enhancing real-time adaptability using reinforcement learning could make the fusion policy more robust and personalized. Most importantly, the present validation was conducted exclusively on a small group of healthy participants (n = 5). Future clinical trials involving stroke survivors and larger cohorts are essential to establish translational impact.

Conclusions

This study presented a novel fatigue-adaptive hybrid EMG–EEG control framework designed to improve user intention detection during upper-limb rehabilitation, specifically targeting elbow movement tasks. The system integrates surface EMG and EEG signals through an adaptive Bayesian fusion model, where signal weighting is dynamically modulated based on a k-NN-based muscle fatigue estimation from EMG features.

For intention detection, SVM classifiers were employed for both EMG and EEG signals, with EEG signals undergoing Common Spatial Pattern (CSP) filtering to enhance discriminative feature extraction. The system was deployed on a lightweight, adjustable elbow rehabilitation robot and evaluated on five healthy participants, demonstrating high accuracy, real-time responsiveness, and robustness under fatigue conditions. Results demonstrated high classification accuracy, robustness under fatigue conditions, and real-time adaptability, thereby validating the system’s feasibility for clinical and home-based rehabilitation applications. These findings establish a foundation for future clinical deployment of fatigue-aware human–robot interaction systems and motivate further exploration into adaptive, multimodal rehabilitation interfaces.

Author contributions

Conceptualization, I.B.A. and Y.B.; Methodology, I.B.A. and Y.B.; Software, I.B.A.; Validation, I.B.A. and Y.B.; Resources, I.B.A. and Y.B.; Writing—original draft preparation, I.B.A. and Y.B.; Supervision, Y.B. and A.A.; Project administration, Y.B. and A.A.; Funding acquisition, A.A.

Funding

The authors extend their appreciation to the King Salman center For Disability Research for funding this work through Research Group no KSRG-2024-360.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.