GRABMyoFlow - Dataset extension

@article{PhysioNet-grabmyo-flow-1.0.0,
  author = {Ehlert, Carl Linus and Tung, James and Pradhan, Ashirbad},
  title = {{GRABMyoFlow - Dataset extension}},
  journal = {{PhysioNet}},
  year = {2026},
  month = apr,
  note = {Version 1.0.0},
  doi = {10.13026/ah52-zq51},
  url = {https://doi.org/10.13026/ah52-zq51}
}

Abstract

This dataset is an extension of the Gesture Recognition and Biometrics Electromyogram (GRABMyo) dataset, a comprehensive collection of surface electromyography (sEMG) signals from the forearm and wrist while performing static gestures, designed for hand gesture recognition (HGR) and biometric authentication research. The study extends the dataset in two key aspects: 1) by adding wrist-only recordings from 20 newly recruited healthy adults, and 2) by including a dynamic trial at the end of each session. The extension study follows the protocol established in the original study to ensure methodological consistency and reliable data acquisition. Combined with prior wrist recordings, the extended dataset now includes data from 63 participants, significantly increasing the sample size. Signal strength and quality in the new recordings have been validated to match the standards of the original dataset, ensuring uniformity across all subjects. The dynamic gesture recordings included in the extension capture the natural transitions between hand poses, better simulating real-world movement scenarios. These dynamic data enable new avenues for continuous authentication and gesture-event segmentation, directly addressing prior limitations of static-only datasets. Data acquisition employed the established electrode configuration, and open access to the data in the waveform database (WFDB)-format is provided to support the widespread use of research. This enriched dataset enhances both the scale and temporal complexity of the GRABMyo dataset, delivering a robust resource for advancing sEMG-based HGR for human-computer interaction or biometric systems research.

Background

Surface electromyography (sEMG) has emerged as a pivotal technology for developing intuitive human-machine interfaces, prosthetic control systems, and biometric authentication methods.

This work builds upon the original GRABMyo dataset [1], a comprehensive collection of forearm and wrist sEMG recordings from 43 participants performing static hand and wrist gestures. While the original dataset established a robust baseline, the translation of sEMG-based applications from controlled laboratory environments to practical real-world deployment continues to face substantial challenges that limit their widespread adoption and clinical utility.

a) Current Small Size and Single-Session Datasets

A persistent limitation in current sEMG research is the reliance on datasets with insufficient participant numbers for robust machine learning model development. Recent comprehensive reviews highlight that most publicly available sEMG datasets contain fewer than 50 participants, with many prominent datasets including as few as 10-20 subjects [2, 3, 4, 5]. This limitation severely constrains the generalization capability of trained models, as evidenced by significant performance degradation when algorithms trained on small cohorts are applied to new users [6, 7, 8]. Cross-participant studies have demonstrated that classification accuracy can drop by 20% to 40% when transitioning from intra-subject to inter-subject scenarios, particularly with datasets containing fewer than 30 participants [6, 7, 9]. The critical need for larger participant pools has intensified as the field advances toward developing universal, population-level models capable of accommodating the substantial inter-individual variability in sEMG signal characteristics across anatomical differences, force variations, and other physiological factors [2, 4, 6].

Complementing the challenge of limited sample sizes, another critical barrier to practical deployment is the predominance of single-session recordings in existing sEMG datasets, which fail to capture the temporal variability and day-to-day changes in signal characteristics that occur in real-world applications [1, 5, 10]. Multi-session data collection is essential for developing robust algorithms that can maintain consistent performance despite electrode displacement, physiological variations, and skin impedance changes that inevitably emerge across multiple recording days [11, 12, 13, 14]. Studies examining cross-session performance have revealed substantial accuracy degradation when models trained on single-day data are applied to recordings from subsequent days, with non-stationary factors including electrode shifts and skin-surface variations seriously affecting system performance [1, 11, 15, 16], highlighting the critical necessity for longitudinal datasets collected across extended timeframes with controlled but realistic electrode repositioning protocols.

b) The Gap in Dynamic sEMG Data

Traditional sEMG research has predominantly focused on static gesture recognition, where participants hold isometric contractions for predetermined durations [10, 17]. While this approach simplifies signal processing and analysis, it fails to capture the rich temporal dynamics inherent in natural human movement and interaction [17, 18]. Dynamic gesture recognition, which involves continuous motion sequences and transitions between poses, represents a more realistic and challenging paradigm that better reflects real-world usage scenarios in prosthetic control, gesture-based interfaces, and human-computer interaction [13, 17]. Recent studies have demonstrated that models trained exclusively on static data exhibit significantly poor performance compared to training on datasets including dynamic recordings, with average recognition error rates improving by 7–11% [13, 17]. Despite this recognized need for comprehensive training approaches, datasets featuring both static and dynamic protocols remain scarce, limiting the development of algorithms capable of handling natural, continuous inputs and transitions required for practical applications [17, 19, 20].

While recent efforts have begun to address some of these limitations through the development of larger datasets and benchmarking initiatives [20], significant gaps remain. Furthermore, existing large-scale datasets focus primarily on forearm recordings, with limited emphasis on wrist configurations that are essential for practical wearable applications [2, 21].

The present dataset extension addresses these critical gaps by providing, firstly, a substantial increase in participant numbers through the addition of 20 new subjects to create a combined cohort of 63 participants, thereby enhancing statistical power and cross-user generalization potential. Secondly, comprehensive dynamic gesture recordings are utilized for the purpose of capturing natural movement transitions and continuous motion sequences. Finally, a multi-session data collection over extended time periods is included to enable robust evaluation of temporal stability and adaptation strategies. As a result, this resource enables systematic exploration of population-level modeling and the development of algorithms capable of bridging the gap between laboratory research and practical deployment in real-world sEMG applications.

Methods

The overall methodology was designed to maintain consistency with the original GRABMyo study [1] while incorporating a new dynamic recording task specific to the extension dataset. A minor adaptation was made to focus sEMG data collection solely on the wrist to reduce setup complexity, shorten session duration, and focus on data most relevant for real-world applications.

A. Participants and Ethics Protocol

The extension of the dataset included a total of 20 healthy participants (12 M, 8 F), with a mean age of 24.25 ± 6.03 years, an average forearm length (measured from the styloid process at the wrist to the olecranon at the elbow) of 23.49 ± 1.44, cm and an average wrist circumference (measured 2 cm away from the ulnar styloid process) of 16.35 ± 1.35 cm. Recruitment was conducted through word of mouth and posters at the University of Waterloo. All participants received an information letter describing the study objectives, protocols, and data use. The procedures were explained verbally with opportunities for questions, and participants were free to withdraw from the study at any point. Prior to enrollment, each participant completed an allergy screening and provided written informed consent. The study conformed to the Declaration of Helsinki and was approved by the Office of Research Ethics at the University of Waterloo (protocol #47560). Table 1 summarizes the measured characteristics of the extension and the combined dataset. If a participant was unable to attend a scheduled session, a window of one day was allowed for the second session, and up to three days for the third session, to improve participation flexibility. Only participants who completed all three recording sessions are included in the dataset. Exclusion criteria were allergies to adhesives, gels, or rubbing alcohol; existing muscle pain; or unexpected personal circumstances that prevented completion of the recording sessions.

B. Experimental Setup

A laptop was used to synchronously control the presentation of gesture and transition prompts on a secondary monitor and the data recordings. The sEMG was acquired using the commercial amplifier (EMGUSB2+ from OT Bioelettronica), configured with a gain of 500 and a sampling rate of 2048 Hz. Pre-gelled monopolar sEMG electrodes (Ambu Neuroline 720-00-S/25) were used. Prior to the experiment, the participant’s right forearm was cleaned with rubbing alcohol, measured, and electrode positions were marked. In cases of excessive hair, the area was shaved to ensure good electrode contact. A total of 14 electrodes were used: two reference electrodes located at the pisiform and olecranon, and twelve recording electrodes arranged in two rings on the wrist/lower forearm. The distal ring was placed 2 cm proximal to the ulnar styloid process, with a center-to-center distance of 2 cm to the proximal ring. To simulate real-world electrode shift, no permanent markings were made on the skin; however, to standardize electrode labeling and positioning between participants, the first electrode of each ring was aligned with the centerline of the elbow crease. The placement of the electrodes is illustrated in the file electrode_placement.png.

C. Session Protocol

The participants were seated comfortably, with their right forearm resting in a neutral position (neither supinated nor pronated) and supported at the level of the chair armrest, ensuring that no electrodes contacted the chair surface.

Static Gesture Recordings: After a brief introduction and a test trial to familiarize participants with the protocol and software interface, static gesture data was collected as in the original study [1]. Visual prompts instructed participants to perform the randomized sequence of hand and wrist gestures at moderate force, each held for 5 s, followed by an 8 s rest period. Sixteen gestures and a rest gesture were recorded in each session. Each participant performed seven repetitions (“runs”) per session; each run comprised all 16 static gestures (trials), followed by the rest pose. Between runs, the quality of the signal and the comfort of the participants were monitored by the experimenters, allowing additional rest as needed. No normalization or maximum voluntary contraction calibration was performed, in accordance with the original protocol. The gesture set is visualized in the file GestureList_static.jpg.

Dynamic Gesture Recordings: Dynamic gestures, defined as transient hand movements between two static gestures, were performed following the static trials. In each session, participants also performed two repetitions (runs) of 30 dynamic transitions (dynamic trials) (GestureList_dynamic.jpg), using the same hand and wrist gesture set as the static protocol. Each 10 s long dynamic trial consisted of the following sequence: (1) hold gesture A for 4 s, (2) transition over 2 s into gesture B, (3) hold gesture B for 4 s followed by a rest of 5 s, resulting in a total duration of 7.5 min per run. The dynamic gestures were randomly paired to cover varied transitions between different hand (gestures 1-10, 15, 16) and wrist gestures (gestures 11-14). The software prompted the participants and recorded timestamps to distinguish between each part of the trials. Failures to follow timing, perform the correct gestures, or excessive noise due to rapid movement prompted a repeat of the affected trial.

D. Signal Processing and Quality Assurance

All recorded sEMG signals were first bandpass filtered between 10 Hz and 500 Hz using a fourth-order Butterworth filter, and a 60 Hz notch filter was applied to remove power line interference, consistent with the original dataset protocol. After each run, the raw signals were reviewed for timing errors, excessive noise, or evidence of incorrect gesture execution. Trials exhibiting these were flagged for correction. Correction recordings were performed at the end of each session to replace or supplement problematic trials. For the final dataset, only the error-corrected recordings are provided. The provided file subject-info.csv enumerates the trials that have been substituted during the error-correction procedure.

Data Description

The extension dataset is structured for immediate access to static and dynamic sEMG recordings, available in the WFDB file format.

A. Data Organization

At the top level, the dataset is divided into static_WFDB and dynamic_WFDB directories. Within each, participant sub-folders are organized by session.

• Static recordings: Provided as individual files for each trial, matching the folder structure and naming conventions of the original GRABMyo dataset. As the provided data are an extension of the data from participants 1 to 20, they are mapped to indices 44 to 63.

• Dynamic recordings: Provided as continuous recordings of each of the 30 trials that include the three different segments: hold gesture A, transition, and hold gesture B for each of the two runs per session.

B. Folder organization

For both datasets, the three top-level folders represent the three sessions. Within each session folder, there are 20 sub-folders for each of the participants. The three folders for the three data collection sessions are: "session1", "session2", and "session3". Each of the folders contains subfolders named "sessioni_subjectj" for each of the participants. Each subject subfolder contains data recording files named "sessioni_subjectj_gesturek_trialt.dat" and "sessioni_subjectj_gesturek_trialt.hea". For the static extension data files, the indexes and ranges are, i ∈ {1,2,3} for the respective session indexes, j ∈ {44,45,...,63} for the respective subject index, k ∈ {1,2,...,17} for the respective gesture indexes and t ∈ {1,2,...,7} for the respective trial indexes. Whereas for the dynamic data files the index ranges are j ∈ {1,2,...,20} for the respective subject index, k ∈ {1,2,...,30} for the respective gesture indexes and t ∈ {1,2} for the respective trial indexes.

There are 16 signals/channels for the EMG recordings, which are sampled at 2048 Hz (10240 x 16). Of these 16 channels, 12 channels are collected from the wrist in a bipolar configuration (6+6 for the wrist in the form of 2 rings, as shown in the electrode_placement.png file), while 4 channels remain unused.

The wrist channels are numbered {W1-W6} (corresponding to the proximal wrist ring) and {W7-W12} (corresponding to the distal wrist ring). The unused channels are listed as {U1-U4} and have been provided to distinguish the rings of electrode setup. The file electrode_placement.png visualizes the exact electrode locations. The sample code provided with the data demonstrates channel selection for both the forearm and wrist configurations; see grabmyoflow_wfdb_to_mat_static.py and grabmyoflow_wfdb_to_mat_dynamic.py. Additional code is provided for visualizing the data in .mat format and for feature extraction using the frequency division technique (FDT); see grabmyoflow_feature_extraction.m [1].

C. File Formats & Additional Files

WFDB files consist of:

• One .dat file containing the raw binary multichannel wrist sEMG signals.
• One .hea header file providing metadata like trial channel numbers or sampling frequency.

Code Files:

• grabmyoflow_wfdb_to_mat_static.py
• grabmyoflow_wfdb_to_mat_dynamic.py
• grabmyoflow_wfdb_visualize_export.py
• grabmyoflow_feature_extraction.m
• grabmyoflow_visualize_static.m
• grabmyoflow_visualize_dynamic.m

Additional Files:

• subject-info.csv: Contains anonymized participant information, session details, and detailed trial-level flags indicating corrections in the static and dynamic data.
• readme.txt: Offers instructions on dataset structure, naming conventions, code usage and requirements.
• GestureList_static.jpg: Provides visual and textual references for the static gestures.
• GestureList_dynamic.jpg: Provides visual references for the dynamic transitions.
• MotionSeq.txt: Provides textual references for the dynamic transitions.
• electrode_placement.png: A visualization of the electrode positions and channel numbers.
• DeviceInfo.pdf: Datasheet for the OT-Biolab USB2+
• segmentEMG.m: Helper function for segmentation
• featiFDTI.m: Helper function for feature extraction

Usage Notes

The data may be used for the following investigations (from GRABMyo [22]).

1. Improving biometric authentication. A multicode EMG-based biometric framework was introduced for a secure biometric system, and different fusion strategies were discussed [23]. A standard biometric authentication system consists of four modules: 1) sensor module that collects the biometric data, 2) feature extractor for generating feature vectors utilized as biometric entries, 3) matcher module that compares with genuine user’s template to generate a score, and 4) decision module to grant access or rejection based on its score threshold. The multiday performance analysis of multicode EMG-based biometrics needs to be investigated.
2. Biometric identification. Another major biometric application is the identification mode, where the system predicts the identity of the presenting user by finding the closest match. As per the definition, the identification is a more error-prone application than the authentication as the system makes N comparisons, where N is the number of users enrolled in the database. Therefore, the factors affecting the system, such as multiple days and sample size of the database, need to be investigated for real-life applications [24].
3. Subject-independent gesture recognition. Extensive research on EMG has been performed on gesture recognition with application in rehabilitation using prosthetic and orthotic devices, home application for assisting daily activities, virtual environment control and sign language recognition [25, 26]. However, with an increase in class labels, there is a training burden for setting up machine learning models [27]. Recent studies have suggested deep learning techniques for cross-user calibration-free which trains generalized models using the population data, and hence reduce the training burden [28, 29]. The presented large-sample dataset can provide adequate resources for such calibration-free models.
4. Electrode shift invariant techniques. One of the significant factors affecting the cross-day performance is the shift in the electrode positions. It is impossible to fix the location of armband electrodes on the forearm and wrist for daily-wear use. These variations affect the performance of both the EMG-based biometric and gesture recognition applications. Some techniques such as classification model adaptation [16] and feature space transformation using transfer learning [30] have been suggested to address the electrode shift variations. These techniques could be further investigated to potentially improve biometrics and gesture recognition performance.

This GRABMyoFlow extension significantly expands the available GRABMyo dataset.

• The inclusion of 20 additional participants enhances the statistical reliability and generalizability of models trained on these data.
• The new dynamic gesture recordings provide sequences capturing transitions and continuous movements, supporting the development of algorithms capable of real-world gesture segmentation and continuous authentication.

The dataset is suitable for a range of analyses, including, but not limited to, Human-Gesture Recognition (HGR) using machine learning approaches, sEMG-based biometrics, and robustness testing against real-world artifacts.

Known Limitations

While this dataset extension offers substantial advancements, researchers should consider certain boundaries inherent to the experimental design and current sEMG applications.

First, data collection was restricted to able-bodied individuals without neuromuscular or dermatological conditions. Consequently, the signal characteristics may not fully generalize to clinical demographics, such as individuals with upper-limb amputations.

Second, the sessions were conducted in a highly controlled laboratory setting where participants remained seated with their arms in a resting posture. This stationary setup minimizes naturally occurring motion artifacts, postural shifts, and baseline fluctuations that would typically affect signal quality during unconstrained daily activities.

Furthermore, the protocol required participants to maintain a consistent, moderate level of exertion, meaning the dataset does not capture the natural variability that occurs when gestures are performed with varying degrees of muscle force.

Finally, although the dataset incorporates multi-session recordings to simulate temporal changes, bridging the gap to continuous, daily-wear applications remains complex. The data relies on a standardized sensor placement protocol that cannot be perfectly replicated by end-users in everyday scenarios. Developing algorithms that remain robust against the inevitable day-to-day displacement of wearable sensors continues to be an ongoing challenge that models utilizing these data must navigate.

Release Notes

Version 1.0.0: Initial release.

Ethics

The experiments were conducted following the Declaration of Helsinki and the research protocol was approved by the Office of Research Ethics of the University of Waterloo (protocol #47560). The authors declare no ethics concerns.

Acknowledgements

This research was conducted with the support of the Neural Rehabilitation Lab (NRE) Lab at the University of Waterloo. The author would like to express sincere gratitude to Professor James Tung, Ashirbad Pradhan, and Seoyeon Woo for their valuable guidance, technical expertise, and assistance with the experimental design and data analysis. The author also thanks the volunteers for their time and participation in this study.

Conflicts of Interest

The authors have no conflicts of interest.

References

1. Pradhan A, He J, Jiang N. Multi-day dataset of forearm and wrist electromyogram for hand gesture recognition and biometrics. Sci Data. 2022;9(1):733
2. Yang Z, Zhou X, Lin R, Ma J, Fang Y. Benchmarking out-of-distribution generalization and adaptation for electromyography. Proc NeurIPS Datasets Benchmarks. 2024
3. Jiang X, Fang Z, Zhu X, Hu Z, Luo M, Yang Y. Open access dataset, toolbox and benchmark processing results of high-density surface electromyogram recordings. IEEE Trans Neural Syst Rehabil Eng. 2021;29:1035-1046.
4. Campbell E, Phinyomark A, Scheme E. A comparison of amputee and able-bodied inter-subject variability in myoelectric control. IEEE Trans Neural Syst Rehabil Eng. 2021.
5. Benatti S, Farella E, Gruppioni E, Benini L. Analysis of robust implementation of an emg pattern recognition system with off-the-shelf components. Biosignals. 2014:45-54.
6. Eddy S, Xiong V, Wang M, et al. Large multi-user models enable robust zero-shot emg classification. bioRxiv. 2024.
7. Campbell E, Phinyomark A, Scheme E. Deep cross-user models reduce the training burden in myoelectric control. Front Neurosci. 2021;15:657958.
8. Fratti R, Marini N, Atzori M, Müller H, Tiengo C, Bassetto F. A multi-scale cnn for transfer learning in semg-based hand gesture recognition for prosthetic devices. Sensors. 2024;24(22):7147.
9. Scheme E, Campbell E, Phinyomark A. Inter-subject variability in myoelectric control. IEEE Trans Biomed Eng. 2020.
10. Kaczmarek P, Mańkowski T, Tomczyński J. Putemg—a surface electromyography hand gesture recognition dataset. Sensors. 2019;19(16):3548.
11. Karrenbach M, Preechayasomboon P, Sauer P, Boe D, Rombokas E. Deep learning and session-specific rapid recalibration for dynamic hand gesture recognition from EMG. Front Bioeng Biotechnol. 2022;10:1034672.
12. Hu Q, Azar GA, Fletcher AK, Rangan S, Atashzar SF. ViT-MDHGR: Cross-day reliability and agility in dynamic hand gesture prediction via hd-semg signal decoding. IEEE J Sel Top Signal Process. 2024;18(3):419-430.
13. Wang S, Wang X, He Y. Analysis of electrode locations on limb condition effect for gesture recognition. Sens Actuators A Phys. 2024. doi:10.1016/j.sna.2024.115177.
14. Botros FS, Phinyomark A, Scheme E. Electromyography-based gesture recognition: Is it time to change focus from the forearm to the wrist? IEEE Trans Industr Inform. 2020;18:174-184.
15. Kyranou I, Christofi V, Hadjipanayi A, Kyriacou E. Emg dataset for gesture recognition with arm translation. Sci Data. 2025. doi:10.1038/s41597-024-04296-8.
16. Vidovic MM, Hwang HJ, Amsüss S, Hahne JM, Farina D, Müller KR. Improving the robustness of myoelectric pattern recognition for upper limb prostheses by covariate shift adaptation. IEEE Trans Neural Syst Rehabil Eng. 2015;24:961-970.
17. Yang Z, Ma J, Fang Y. Dynamic gesture recognition using surface emg signals with recurrent neural networks. Front Bioeng Biotechnol. 2021;9:779353.
18. Gigli A, Gijsberts A, Castellini C. A comparison between static and dynamic data acquisition. In: 2019 IEEE 16th International Conference on Rehabilitation Robotics (ICORR). IEEE; 2019.
19. Tacca N, Fratti R, Atzori M, De Michieli L, Muceli S. Wearable high-density emg sleeve for complex hand gesture recognition. Sci Rep. 2024;14:18543.
20. Salter S, O'Malley B, Mignone S, Zukowski B, He J. Emg2pose: A large and diverse benchmark for surface electromyographic hand pose estimation. In: Proceedings of the 38th International Conference on Neural Information Processing Systems (NIPS '24). Curran Associates Inc; 2025.
21. Ni S, Al-qaness MA, Hawbani A, Al-Alimi D, Abd Elaziz M, Ewees AA. A survey on hand gesture recognition based on surface electromyography: Fundamentals, methods, applications, challenges and future trends. Appl Soft Comput. 2024;166:112235.
22. Jiang N, Pradhan A, He J. Gesture Recognition and Biometrics ElectroMyogram (GRABMyo) (version 1.1.0). PhysioNet; 2024. doi:10.13026/89dm-f662.
23. Pradhan A, He J, Jiang N. Score, Rank, and Decision-Level Fusion Strategies of Multicode Electromyogram-based Verification and Identification Biometrics. IEEE J Biomed Health Inform. 2021.
24. Pradhan A, He J, Jiang N. Performance Optimization of Surface Electromyography based Biometric Sensing System for both Verification and Identification. IEEE Sens J. 2021.
25. Dwivedi A, Kwon Y, Liarokapis M. EMG-Based Decoding of Manipulation Motions in Virtual Reality: Towards Immersive Interfaces. In: 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE; 2020:3296-3303.
26. Wu J, Sun L, Jafari R. A wearable system for recognizing American sign language in real-time using IMU and surface EMG sensors. IEEE J Biomed Health Inform. 2016;20(5):1281-1290.
27. Marano G, Brambilla C, Mira RM, Scano A, Müller H, Atzori M. Questioning Domain Adaptation in Myoelectric Hand Prostheses Control: An Inter-and Intra-Subject Study. Sensors. 2021;21(22):7500.
28. Côté-Allard U, et al. Unsupervised domain adversarial self-calibration for electromyography-based gesture recognition. IEEE Access. 2020;8:177941-177955.
29. Côté-Allard U, Campbell E, Phinyomark A, Laviolette F, Gosselin B, Scheme E. Interpreting deep learning features for myoelectric control: A comparison with handcrafted features. Front Bioeng Biotechnol. 2020;8:158.
30. Prahm C, et al. Counteracting electrode shifts in upper-limb prosthesis control via transfer learning. IEEE Trans Neural Syst Rehabil Eng. 2019;27(5):956-962.