Muscle activation detection with surface electromyography: A combined signal processing and machine learning approach

Abstract

Surface electromyography (sEMG) is a noninvasive method for monitoring muscle activity, essential in rehabilitation, prosthetics, sports, and medicine. However, current sEMG systems struggle with signal noise, unclear muscle activation detection, and limited adaptability. This study proposes a method combining signal processing and machine learning to enhance muscle activation detection. Experiments on EMG data from muscle contractions (n = 140 per test, with 28 participants, five biceps curl repetitions, five wrist curl repetitions each) demonstrated 92.23% overall accuracy (3.83% false positives, 3.94% false negatives) for biceps tests and 91.11% overall accuracy (4.36% false positives, 4.53% false negatives) for wrist tests, and an average recall of 96% significantly outperforming traditional methods. This approach highlights potential applications in real-world biomedical settings from rehabilitation and prosthetics to sports science.

Keywords

Surface electromyography muscle activation biomechanics signal processing machine learning rehabilitation

Introduction

Surface electromyography (sEMG) is a noninvasive measurement technique that records muscle-generated electrical signals by two or three electrodes placed on the skin overlying the muscle of interest. These electrodes measure the electrical voltage difference generated by muscle fibers in different states. This raw signal, known as electromyogram, is faint and mixed with noise from various sources; as such, it needs careful amplification and filtering to be correctly read. This reading can be analyzed to provide real-time feedback on the timing and intensity of muscle contraction, making it extremely valuable for applications in sport science, biomechanics, rehabilitation, and assistive technology.^1,2 Medical applications range from diagnosing neuromuscular disorders like muscular dystrophy and amyotrophic lateral sclerosis³ to monitoring muscle function and fatigue when recovering from strokes or performing physical therapy.⁴ In sports, EMG helps collecting large datasets to better understand the muscular effort during activities, optimizing training and techniques for maximum efficiency and power.⁵ In augmented reality and robotic applications, they can be used as human–machine interface to capture and classify human motion.^2,5,-7

Despite its advantages, sEMG systems often suffer from signal noise and limited adaptability to physiological variations.^5,8 A variety of hardware solutions and postprocessing methods have been suggested to improve data quality, which partially solve these issues.^1,9 However, the low magnitude of the signal still makes it extremely sensitive to noise and interference. A common solution observed in literature includes the usage of multiple EMG sensors or fusion with other types of sensors, in order to extract either force,¹⁰ motion,^11,12 or personal¹³ information from the acquired data. This option, however, involves complex systems and introduces additional data and processing that is not always necessary when the EMG is used to identify only muscle activation, for example, to drive a prosthesis. Artificial intelligence (AI) has often been used in multisensor scenarios, with previous works relying on sensor fusion to perform pattern recognition^14,15; other previous works adopt AI to analyze sEMG datasets to evaluate overall muscle activity or performance with a single global performance parameter.¹⁶ Recent works generally rely on multiple sEMG sensors to identify muscle activation pattern throughout related muscles,^17,18 sometimes combined with inertial measurement units.¹⁹

Overall, there is currently a research gap in the literature regarding the identification of muscle patterns without the use of multiple sensors (either all sEMG or with a combination of sEMG and other sensing technologies), leading to increased cost and complexity creating a barrier to the widespread adoption of sEMG in daily task or sport monitoring. This study aims at this gap, providing a streamlined method for muscle activation detection with an individual sEMG sensor. Here we show how muscle activation patterns can be detected with a single sEMG, enabling simpler designs for sEMG-activated controllers (e.g. in prostheses) by integrating signal processing techniques—such as filtering, rectification, root mean square (RMS) envelope extraction, and thresholding—with machine learning models (Random Forest, XGBoost) to improve accuracy and reliability.

The article first introduces the three-electrode sEMG sensor used in this study, explaining our experimental protocol over two different upper-limb exercises and our postprocessing pipeline. Then, we report study results, discussing both analytical thresholds and learning-based methods for pattern detection. Finally, we compare the performance of the proposed methods, highlighting the advantages and disadvantages of both.

Material and methods

EMG sensor

This study analyzed a signal obtained through a Beyond EMG sensor (Sensor Medica srl). As a two-channel wearable EMG sensor, it can measure electrical signals from two different muscles at the same time in either a two-electrode (Figure 1(a)) or three-electrode (Figure 1(b)) configuration. Two of the electrodes collect muscle signals, while the optional third electrode, placed on a bone, provides a stable reference for a cleaner signal. Signal strength (gain) can be set to four different levels: 330, 466, 658, and 1170 V/V. The sensor filters signals through a low-pass filter set at 250 Hz and a high-pass filter at 10 Hz to reduce unwanted noise. Noise levels of less than 2 µV RMS ensure clean signals, while its high input impedance of over 100 million ohms helps protect the signal from interference. It can detect signal levels up to 10 mV peak-to-peak (mVpp), which is suitable for capturing muscle electrical activity. The sensor uses a 12-bit analog-to-digital converter to convert the signals into digital form for analysis with a dynamic range of 3.3 V, and users can choose between two sampling rates: 1000 Hz or 500 Hz.

Figure 1.

Beyond EMG sensor (Sensor Medica srl): (a) Two-electrode layout, with two pairs (positive–negative) of electrodes; (b) three-electrode layout with two sets (positive–negative–reference) of electrodes; (c) user interface of the commercial integrated software with raw signal (white) and postprocessed signal with activation (blue) and deactivation (red) detection; (d) detail from an example of inaccurate detection with false negatives (in red despite clear muscle activity in the raw signal).

The software provided with the sensor acquires sensor data and processes them, providing as output the signal plot in the example graphical user interface in Figure 1(c) and automatically detecting muscle activation and deactivation at threshold increase or decrease rates of the acquired values. However, while this code proves fairly reliable in controlled environments and for main muscle groups (e.g. biceps), high rates of false positives (i.e. identifying the muscle as active when it is not) and especially false negatives (i.e. identifying no activation when the muscle is active, visible in the examples in Figure 1(d)) are observed in realistic environments and smaller muscle groups (e.g. wrist), leading to a very inaccurate detection in several use cases (less than 20% accuracy for some exercises). The lack of medical and technical standards on EMG signal acquisition and postprocessing, together with the low accuracy of state-of-the-art muscle activation algorithms, is the main motivation behind this study, which aims to improve EMG signal analysis through new analytical methods and learning-based algorithms.

Experimental protocol

To obtain functional signals with clear muscle activation and deactivation to be used as ground truth, a timed weight-lifting exercise has been performed to generate a sample signal with five repetitions of biceps curls and then five repetitions of wrist curl exercises, repeated for 28 users (n = 140 for biceps; n = 140 for wrist). Figure 2 shows the experimental layout with the two different phases in our experimental acquisitions. The first phase is to acquire biceps activation during a 50 s cycle of biceps curl (5 s rest, 5 s activity) with a 4 kg dumbbell. In the second phase, new electrodes are placed on the forearm to perform an equally timed exercise of wrist curl with the same dumbbell. The latter task, after an empirical tuning, has shown the best results in activating the flexor digitorum superficialis muscle. In both exercises, the skin was precleaned with alcohol, the interelectrode distance was set to 1–2 cm, and placement strictly followed SENIAM guidelines for reproducibility.²⁰ More details are reported in the experimental protocol in Figure 3. This step-by-step procedure has been applied to all the test subjects. When instabilities were observed (e.g. interference, unexpected sensor behavior), the test was discarded and repeated from the beginning.

Figure 2.

Experimental setup with reference electrode positions: on the left, weight-lifting routine for biceps curl; on the right, the EMG sensor on the forearm measures the flexor digitorum superficialis activation in the wrist curl. The blue electrode is the ground (positioned in both cases on the elbow). Red and black electrodes take the signal differences with respect to our body reference; we have positioned the electrodes along the muscle 2 cm of distance each other.

Figure 3.

Experimental protocol for the acquisition of muscle activity during a biceps curl exercise first and a wrist curl exercise later; each exercise is repeated five times during the routine.

Postprocessing

The acquired EMG data in this work have been analyzed through the streamlined postprocessing pipeline here presented. First, an initial calibration is performed asking the user to relax the muscle for 5 s. Then, the acquired data is used to remove the baseline offset (constant background value) from the data, averaging the initial n data points and subtracting this average $x_{0}$ from the rest of the acquired data points $x_{i}$ :

x_{i}^{offset} = x_{i} - x_{0}; x_{0} = \frac{1}{n} \sum_{i = 1}^{n} x_{i}

(1)

After baseline correction to remove constant offsets, rectification was used to convert values to positive as

x_{i}^{r e c t} = | x_{i}^{o f f s e t} |

(2)

The value after equations (1)–(2) provides a cleaner visualization of the raw signal (in blue in Figure 4). Then, the signal has been smoothed by computing its RMS over a moving window. This mobile average, used as in related literature to minimize the effect of outliers on further analysis, represents muscle activation as

x_{i}^{R M S} = \sqrt{\frac{1}{w} \sum_{j = i - w + 1}^{i} {(x_{\dot{j}}^{r e c t})}^{2}}

(3)

where w is the size of the moving window. Higher values of w result in smoother visualization and reduction of chattering but progressively lose information on the signal as well as introducing a delay. Different values of window size have been tested, with good results observed for a 50-sample moving window, as in the RMS envelope example in red in Figure 4.

Figure 4.

Example of EMG signal processing on three repetitions of an example biceps curl exercise, with raw signal in black, processed signal with equations (1)–(2) in blue, and RMS in red.

The challenge of detecting muscle activation has been discussed here with two possible solutions: an analytical one, focusing on the change of the RMS envelope through thresholds in both absolute value and derivative, and a learning-based one, training a neural network to automatically detect activation and deactivation, as described in the two following subsections, respectively.

Analytical detection of muscle activation

The analytical approach for muscle activation aims at determining functional thresholds for both signal absolute value $x_{i}^{R M S}$ and its rate of change ${\dot{x}}_{i}^{R M S}$ , defined numerically as

{\dot{x}}_{i}^{R M S} = \frac{x_{i}^{RMS} - x_{i - 1}^{RMS}}{Δ t},

(4)

where

Δ t

is the sampling time of the acquisition. These thresholds in value

x_{t}^{R M S}

and rate of change

{\dot{x}}_{i}^{R M S}

can then be used to identify muscle activation as a binary state

α_{i}

, equal to 1 when the muscle is active and 0 when it is not. This condition can be written for the absolute value threshold as

{\begin{matrix} x_{i}^{RMS} \geq x_{t}^{RMS} \to α_{i} = 1 \\ x_{i}^{RMS} < x_{t}^{RMS} \to α_{i} = 0 \end{matrix}

(5)

or, alternatively, for the rate of change as

{\begin{matrix} {\dot{x}}_{i}^{RMS} \geq {\dot{x}}_{t}^{RMS} \to α_{i} = 1 \\ {\dot{x}}_{i}^{RMS} \leq {\dot{x}}_{t}^{RMS} \to α_{i} = 0 \\ {\dot{x}}_{t}^{RMS} < {\dot{x}}_{i}^{RMS} < {\dot{x}}_{t}^{RMS} \to α_{i} = α_{i - 1} \end{matrix}

(6)

AI-based detection of muscle activation

To further improve muscle detection accuracy, machine learning methods were integrated in the signal analysis pipeline, using Random Forest and XGBoost classifiers. Preprocessing steps included baseline correction to remove offsets, as in equation (1), followed by a Wiener and Butterworth filter to reduce noise by isolating the frequency range of interest (15–450 Hz) and rectification as per equation (2). Key features capturing time-domain (RMS, variance, waveform length), frequency-domain (median frequency, peak frequency, spectral entropy), and energy-based (Teager-Kaiser Energy Operator and the Hilbert envelope) characteristics were extracted from the filtered signal and used to train a machine learning model to classify muscle activity as active ( $α_{i} = 1$ ) or resting ( $α_{i} = 0$ ); training samples have been checked and classified manually. To address class imbalance in the dataset, Synthetic Minority Over-sampling Technique was applied during model training, ensuring balanced representation of muscle states.

Results and discussion

Results: Analytical detection

Different values of moving window sample size and thresholds have been tested, both in absolute value (mV) and change rate (mV/s), with optimal results obtained for a 50-sample moving window and a threshold value of 0.095 mV/s on change rate, as shown in Figure 5(d). The other tests highlight the general errors observed in EMG sensor analysis, with false positives in Figure 5(a) and false negatives in Figure 5(c).

Figure 5.

Muscle activation identified with analytical methods on three repetitions of an example biceps curl exercise, using different threshold values at signal rate of change to evaluate sensitivity to threshold parameters: (a) threshold at 0.01 mV/s; (b) threshold at 0.10 mV/s; (c) threshold at 1.00 mV/s; (d) threshold at 0.095 mV/s.

Results: AI-based detection

The results of the tests with the AI-based detection method are reported in Table 1, with some examples of acquisition and post-processing reported in Figures 6 and 7, respectively. The results in the table report, for each test subject, the time intervals in which the sensor detects muscle activation ( $α = 1)$ , with an accuracy computed as the ratio between the time in which the muscle was active and the measured activation time, thus evaluating the number of false positives in the sample. Additionally, the time interval with false negatives is also reported, describing the amount of time when the sensor should have detected activation but did not.

Figure 6.

Examples of muscle activation identified with AI-based methods from biceps curls: (a) an acquisition example; (b) an example of an acquisition with increased noise.

Figure 7.

Examples of muscle activation identified with AI-based methods from wrist curls: (a) an acquisition example; (b) an example of an acquisition with a weak signal.

Table 1.

Summary of results, with 92.23% overall accuracy (3.83% false positives, 3.94% false negatives) for biceps tests and 91.11% overall accuracy (4.36% false positives, 4.53% false negatives) for wrist tests.

Test number [–]	Accuracy at $α = 1$ [%]		Time with $α = 1$ [s]		Time with $α = 0$ [s]		Total accuracy [%]
Test type	Biceps	Wrist	Biceps	Wrist	Biceps	Wrist	Biceps	Wrist
1	99.60	94.38	12.54	13.93	0.05	0.82	99.20	89.14
2	96.26	95.74	13.58	13.95	0.61	0.63	92.12	91.60
3	95.34	95.94	13.92	14.39	0.62	0.60	91.27	92.10
4	95.90	96.19	13.84	13.98	0.61	0.54	91.85	92.61
5	97.93	96.63	13.68	13.88	0.32	0.55	95.69	92.95
6	98.20	95.12	13.31	13.07	0.26	0.65	96.32	90.61
7	97.42	96.98	15.75	13.98	0.43	0.44	94.83	94.02
8	94.70	96.41	13.78	13.36	0.86	0.50	89.14	92.93
9	95.38	96.38	13.87	14.23	0.64	0.54	91.17	92.86
10	95.82	94.06	12.88	13.30	0.53	0.84	92.03	88.47
11	95.67	95.50	12.76	12.65	0.62	0.67	91.24	90.70
12	95.13	96.92	13.13	12.71	0.71	0.45	90.25	93.61
13	94.55	94.51	13.31	12.77	0.76	0.73	89.44	89.40
14	96.48	94.57	13.73	12.64	0.49	0.69	93.16	89.67
15	95.61	95.39	12.70	12.78	0.51	0.61	91.92	91.04
16	95.25	95.59	12.70	12.68	0.55	0.62	91.30	91.13
17	95.30	93.62	12.58	13.76	0.65	0.92	90.62	87.75
18	97.42	93.56	13.82	13.14	0.35	0.90	95.01	87.56
19	95.25	95.27	13.13	12.74	0.53	0.63	91.55	90.78
20	98.57	95.67	11.77	13.97	0.19	0.63	97.00	91.54
21	96.75	95.14	13.15	12.72	0.45	0.63	93.55	90.65
22	94.88	94.72	13.14	13.18	0.72	0.74	89.95	89.68
23	93.78	94.92	12.49	12.72	0.81	0.65	88.07	90.31
24	97.42	94.99	12.84	14.10	0.36	0.74	94.76	90.25
25	95.25	95.04	14.15	12.86	0.73	0.46	90.58	91.76
26	95.77	97.94	13.48	12.66	0.62	0.24	91.56	96.12
27	95.44	95.88	13.76	12.29	0.71	0.58	90.76	91.56
28	93.36	95.10	13.08	12.92	0.62	0.64	89.13	90.61

Accuracy is computed as the ratio between the time in which the muscle was active and the measured activation time.

Discussion

Overall, the analytical detection results proved to be sample-dependent, as different samples and acquisitions show different optimal values of thresholds. This analytical threshold-based method provided a simple and reliable approach for identifying muscle activation without relying on machine learning but requires time- and effort-intensive manual calibration for each set of samples. Further, its effectiveness diminishes under high-noise or variable physiological conditions, suggesting the potential benefit of integrating AI techniques for further improvement. Finally, the moving window introduces delays in activation time that reduce overall sensor effectiveness, especially in online usage (e.g. exoskeleton control). For these reasons, the analytical detection method has not been used to post-process all the acquired data of the dataset, given its already low performance on a testing subsample (detection accuracy lower than 70%).

The results for the AI-based detection show different kinds of acquisitions. In Figure 6(a), a typical biceps acquisition is reported: the raw signal, in gray at the top, already shows clearly when the muscle is active and when it is not; after postprocessing, the proposed algorithms correctly detects the first, second, and fifth repetition; in the third and fourth repetition, limited intervals with lower muscle activation are incorrectly labeled as periods of rest (see zeros in the diagram at the bottom). Figure 6(b) shows a much noisier acquisition, with noise partially covering the periods of rest; while this noise would have caused the analytical methods to fail, the proposed algorithm manages to detect activation with good accuracy. In Figure 7(a), the code correctly identifies the five main repetitions, even though some false positives and negatives are observed at the beginning and end of each period of activation, as the detected events are not fully synchronized with the experiments. Figure 7(b) shows how the forearm electrodes acquire a much weaker signal when compared to the bicep layout; this reduced amplitude led to misdetection for the analytical methods, but the AI-based algorithm proved to be robust even in this case, despite a significant noise before the first activation lowering performance.

The main challenge observed during all experiments is the presence of excessive noise, likely caused by electrical interference as it was observed as worse in the presence of electronics (e.g. smartwatch and exoskeletons). This noise makes it difficult to clearly capture muscle activity, leading to unreliable data when assessing muscle strength and timing. Initial tests conducted without the use of weights showed unclear results. Signal amplification solves this problem only partially, as it amplifies the above-mentioned noise in addition to the signal; introducing weights during the tests improved the clarity of the results by forcing higher muscle activation, making the signal more readable. This amplification is particularly needed in the wrist test: biceps signal is generally clear and stable, given the larger size of the muscle, whereas wrist flexors are smaller and resulted in worse signal quality, also because electrodes of the same size were used for both wrist and biceps acquisitions. This signal difference is critical with analytical thresholding, but learning-based methods demonstrate their robustness with their small loss of performance between the different tests (approximately 1.12%).

The analytical method works from a theoretical perspective but requires manual calibration case-by-case, sometimes resulting in low accuracy (less than 70%) even between consecutive tests with the same person. This level of effort represents a critical barrier to the accessibility of sEMG technology. Conversely, the trained models demonstrated strong performance, achieving detection with 92.23% overall accuracy (3.83% false positives, 3.94% false negatives) for biceps tests and 91.11% overall accuracy (4.36% false positives, 4.53% false negatives) for wrist tests, and an average recall of 96%. As such, the AI-based method outperforms threshold-based methods, easily adapting to noisy signals (e.g. Figure 6(b)) as well as weak and variable signals (e.g. Figure 7(b)).

Conclusions

This study developed a framework for analyzing sEMG signals to improve muscle activation detection. By integrating baseline correction, rectification, RMS envelope extraction, and thresholding, the signal processing approach reduced noise and highlighted activation patterns. However, traditional threshold-based methods can struggle with adaptability in complex or noisy conditions (accuracy lower than 70%). To overcome these limitations, machine learning models such as Random Forest and XGBoost were introduced. By using key time-domain, frequency-domain, and energy-based features, these AI-based methods significantly enhanced classification accuracy and robustness. The combined approach achieved high performance levels, offering a reliable method for muscle activation analysis with 92.23% overall accuracy (3.83% false positives, 3.94% false negatives) for biceps tests and 91.11% overall accuracy (4.36% false positives, 4.53% false negatives) for wrist tests. While multiple sensors can still provide more nuanced information on motion and biomechanics, the proposed method shows how a single sensor can still be effective in identifying muscle activation patterns for a simple but efficient analysis as well as potential for real-time exoskeleton control. When compared with previous works, where muscle activation detection is obtained manually during postprocessing or discussed without any quality metrics, these findings provide a simple but effective automatic method, contributing to the integration of sEMG in real-time monitoring or control for rehabilitation, sports, exoskeletons, and prosthetics. Future research could extend this work to additional muscle groups and explore real-time, wireless solutions to improve practical usability in clinical and biomechanical settings.

Footnotes

ORCID iDs

Alessandro Perini

Luca Molinaro

Matteo Russo

Consent to participate

Received written informed consent from participants. This study protocol was reviewed and approved by the Institutional Ethics Committee of the Tor Vergata University Hospital (Policlinico Tor Vergata), Rome, on 21 March 2024 (Approval No. RS. 26.24 CET2 UTV) in accordance with the Declaration of Helsinki.

Funding

This study was partly funded by the Italian Government through the ASSIST project (Grant No. P2022A4ELB, PRIN-PNRR 2022) and Alessandro Perini's DM630 PhD Scholarship, and by Sensor Medica srl.

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

Collected data are reported in the body of the article.

References

Rodríguez-Tapia

Soto

Martínez

, et al. Myoelectric interfaces and related applications: current state of EMG signal processing—a systematic review. IEEE Access 2020; 8: 177999–178021.

Zheng

Crouch

Eggleston

. Surface electromyography as a natural human–machine interface: a review. IEEE Sensors J 2022; 22: 9198–9212.

Gonzalez-Lopez

Gomez-Alanis

Martín Doñas

, et al. Silent speech interfaces for speech restoration: a review. IEEE Access 2020; 8: 177999–178021.

Munoz-Novoa

Kristoffersen

Sunnerhagen

, et al. Upper limb stroke rehabilitation using surface electromyography: a systematic review and meta-analysis. Front Hum Neurosci 2022; 16: 897870.

Herrera-Luna

Rechy-Ramirez

Rios

, et al. Interactive wearable systems to monitor movement and posture during upper body rehabilitation. IEEE Sensors J 2019; 19: 3590–3603.

Xiong

. Advances and disturbances in sEMG-based intentions and movements recognition: a review. IEEE Sensors J 2021; 21: 13019–13035.

Tang

Liu

, et al. Hand motion classification using a multi-channel surface electromyography sensor. Sensors 2012; 12: 1130–1147.

Campanini

Disselhorst-Klug

Rymer

, et al. Surface EMG in clinical assessment and neurorehabilitation: barriers limiting its use. Front Neurol 2020; 11: 34.

Chen

Liu

, et al. Hybrid methods for muscle artifact removal in EEG signals. IEEE Sensors J 2019; 19: 3590–3603.

10.

Disselhorst-Klug

Schmitz-Rode

Rau

. Surface electromyography and muscle force: limits in sEMG–force relationship and new approaches for applications. Clin Biomech 2009; 24: 225–235.

11.

Gohel

Mehendale

. Review on electromyography signal acquisition and processing. Biophys Rev 2020; 12: 1361–1367.

12.

Leelakittisin

Wilaiprasitporn

Sudhawiyangkul

. Compact CNN for rapid inter-day hand gesture recognition and person identification from sEMG. IEEE Sensors J 2021; 21: 13019–13035.

13.

Jiang

Liu

, et al. High-density surface electromyogram-based biometrics for personal identification. Proceedings of the 42nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) 2020; 1: 728–731.

14.

Kim

Shin

Kwon

, et al. EMG-based dynamic hand gesture recognition using edge AI for human–robot interaction. Electronics (Basel) 2023; 12: 1541.

15.

Park

Lee

. EMG pattern recognition based on artificial intelligence techniques. IEEE Trans Rehabil Eng 1998; 6: 400–405.

16.

Gorenshtein

Weisblat

Khateb

, et al. AI-based EMG reporting: a randomized controlled trial. J Neurol 2025; 272: 1–12.

17.

Weber

Goetschi

Zipser

. Modification of the serratus anterior strength test to reduce confounding muscle activation: an EMG-based validation study. Phys Ther Sport 2025; 75: 128–134.

18.

Liu

Song

, et al. Factors, characteristics and influences of the changes of muscle activation patterns for patients with knee osteoarthritis: a review. J Orthop Surg Res 2025; 20: 12.

19.

Bai

Song

, et al. Sliding-window CNN+ channel-time attention transformer network trained with inertial measurement units and surface electromyography data for the prediction of muscle activation and motion dynamics leveraging IMU-only wearables for home-based shoulder rehabilitation. Sensors 2025; 25: 1275.

20.

Seniam Guidelines. Available online at http://www.seniam.org/, last accessed 30/10/2025.