---
title: "Deep Learning 1D-CNN Accurately Classifies Nine Finger Movements from Surface EMG Signals"
id: "plos-one-8-deep-learning-driven-fine-resolution-finger-movement-recognition-from-surface"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-8-deep-learning-driven-fine-resolution-finger-movement-recognition-from-surface"
content_type: "clinical_feed_article"
specialty: "Neurology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361050"
published_at: "2026-10-22T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Deep Learning 1D-CNN Accurately Classifies Nine Finger Movements from Surface EMG Signals
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-8-deep-learning-driven-fine-resolution-finger-movement-recognition-from-surface
- **Specialty:** [Neurology](https://medichelpline.com/clinical-feed/neurology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361050)
- **Published At:** 2026-10-22T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- Investigators used the publicly available Ninapro DB2 dataset featuring 40 subjects performing nine distinct finger movements to train a custom 1D convolutional neural network (CNN). - The study targeted fine-resolution finger motion recognition rather than broad hand gestures to enhance prosthetic control precision. - Data preprocessing included band-pass and notch filtering, wavelet-based denoising, full-wave rectification, and per-channel z-score normalization. - Windowing of EMG signals used 400 ms windows with 50% overlap to capture temporal motor patterns effectively. - The 1D-CNN was trained within-subject using repetitions 2 and 5 to minimize temporal data correlation and prevent memorization bias. - Classification accuracy reached 90.1% overall for the nine finger movements, with detailed analysis of sensitivity, specificity, precision, and F1 score per movement. - The study presented confusion matrices and ROC curves highlighting model strengths and areas for misclassification. - The model architecture focused on computational efficiency suitable for potential integration into low-power prosthetic devices. - Findings emphasize the gap in prior literature that focused on overall accuracy with limited per-movement diagnostics. - Authors recommended future exploration of window size optimization and potential transformer architectures balancing accuracy and computational demand. - The work advances EMG decoding granularity critical for restoring natural dexterity in upper-limb prosthetic users and informs design of refined control algorithms.
## Clinical Analysis & Structured Key Points
Deep learning–driven fine-resolution finger movement recognition from surface EMG signals | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Surface electromyography (sEMG) has emerged as an impactful tool for controlling advanced prosthetic devices by detecting and recording muscle electrical activity during motion. This study presents a unique technique for recognizing nine distinct classes of finger movements by utilizing the Ninapro database. To address an underexplored domain of upper-limb prosthetic control, this research targets fine-resolution individual finger movements rather than the broad range of hand gestures. A custom-made one-dimensional convolutional neural network (1D–CNN) was developed and validated using a within-subject evaluation strategy and signals. The proposed model achieved an overall classification accuracy of 90.1%. An in-depth sensitivity/specificity analysis for each finger movement class was administered to gain key insights into current recognition capabilities. The findings presented in this study demonstrate the viability of fine-resolution finger-movement recognition, offering foundational insights that may inform the future development and refinement of control systems in the prosthetic field. Citation: Alquran H, Alawneh H, Alsalatie M, Khader A, Mustafa WA, Alawneh KZ (2026) Deep learning–driven fine-resolution finger movement recognition from surface EMG signals. PLoS One 21(10): e0361050. https://doi.org/10.1371/journal.pone.0361050 Editor: Alessandro Mengarelli, Universita Politecnica delle Marche Facolta di Ingegneria, ITALY Received: January 3, 2026; Accepted: October 5, 2026; Published: October 22, 2026 Copyright: © 2026 Alquran et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The data used in this study are publicly available from the Ninapro Repository (DB2) at http://ninapro.hevs.ch . The authors had no special access privileges and obtained the data through the same mechanism available to other researchers. Funding: The author(s) received no specific funding for this work. Competing interests: no competing interests exist. 1. Introduction Current advancements in prosthetic technologies reserve the potential to improve upper-limb rehabilitation by offering enhanced motor functionality via electromyography (EMG)-controlled devices. The EMG-controlled devices are designed to restore a more natural and dexterous sense of hand function for individuals characterized by upper limb loss. These devices therefore represent a critically needed advancement in assistive technology for this population [ 1 , 2 ]. Constrained by both mechanical simplicity and a primitive control paradigm, traditional prosthetic devices often fail to replicate complex human motion [ 3 ]. EMG signals, designed to reflect the electrical activity of muscles, offer valuable insights for replicating natural hand gestures. However, their application is challenged by the accurate decoding of motor intent due to their non-stationary nature and reported susceptibility to noise from several factors, including muscle fatigue, electrode displacement, and intersubject variability [ 4 ]. The rise of artificial intelligence (AI) and machine learning (ML) applications has brought forth an intention to facilitate a paradigmatic shift within EMG-based control systems [ 5 – 7 ]. Deep learning models have demonstrated high proficiency in detecting complex spatiotemporal patterns, facilitating the automatic extraction of discriminative features from raw input data [ 8 – 10 ]. High accuracy, despite signal noise and inter-user variability, is observed in applications of convolutional neural networks (CNNs) and recurrent neural networks (RNNs), thereby propelling recognition systems forward [ 11 , 12 ]. While CNNs have been extensively investigated for sEMG-based pattern recognition, more recent architectures, namely Vision Transformers (ViTs), have also shown promising results in related classification tasks. The potential advantages of utilizing ViTs include capturing both long-range dependencies and global context within signals, albeit often at the cost of higher computational demands than CNNs. For instance, transformer models originally designed for image recognition [ 13 ] have been adapted for high-density surface EMG gesture recognition, achieving accuracy that challenges or exceeds CNN benchmarks [ 14 ]. However, their large parameter counts and computational needs can limit their use in low-power prosthetic hardware, which led to the adoption of a compact 1D-CNN in this study; nevertheless, transformer-based methods remain an exciting avenue for future research. A study employing the Hilbert-Huang Transform and deep neural networks achieved up to 98.5% accuracy, depending on the hand movement analyzed [ 6 ]. Naser et al. proposed a semi-supervised learning autoencoder framework that integrates a multilayer neural network for sEMG-based classification of five basic hand gestures, reporting an accuracy of 99.26%. However, the analysis was limited to simple hand gestures [ 7 ]. In another study, the GrabMyo dataset was utilized with ML algorithms applied for gesture classification, where a peak accuracy of 90.6% was achieved with a linear discriminant analysis-support vector machine combination [ 15 ]. Furthermore, Kim et al. developed an artificial neural network-based system capable of classifying seven hand gestures and three grasp types, reporting an accuracy of 97.49% but requiring extensive training sessions, which may have potentially skewed the reported outcomes [ 16 ]. As such, the persistent misclassification of hand gestures emphasizes the need to refine feature extraction techniques. In the context of this study, the Ninapro dataset emerged as a promising tool that provides a comprehensive multi-channel EMG repository under standardized conditions [ 3 , 17 ], along with broad coverage of diverse hand gestures, which helps facilitate benchmarking and improve overall control algorithms [ 3 , 18 ]. Atzori et al. used a parallel-windows feature extractor to extract signal features from the Ninapro DB1–5 datasets, achieving sub-80% classification accuracy with a Random Forest classifier [ 19 ]. Using Ninapro DB5, Wang et al. combined CNNs, long short-term memory (LSTM), and convolutional block attention modules to achieve a 92.15% accuracy in hand gesture recognition [ 20 ]. Emima et al. proposed a multi-scale 1D–CNN framework, where accuracies of 93.4%, 92.8%, 91.3%, and 94.1% across Ninapro DB1, DB2, DB5, and DB7 were achieved respectively. Nonetheless, the computational complexity of the presented models poses significant challenges to future real-time deployment in low-power prosthetic devices. Despite the progress made in sEMG-based hand gesture recognition, several challenges remain. The common presence of both noise and artifacts in sEMG signals poses significant obstacles to achieving reliable and validated classification results. In addition, generalizing models across varying subjects and environments remains a critical issue. Despite deep learning models such as CNNs and RNNs achieving high performance on specific datasets, performance commonly degrades when applied to new individuals or under varying conditions. Furthermore, an ongoing challenge persists in the trade-off between model complexity and interpretability, particularly in the context of medical applications, where transparency plays a key role in decision-making. Previous studies have primarily focused on hand-gesture movements without specifically addressing individual finger movements, which are fundamental to hand gestures. From a clinical perspective, reliably decoding individual finger movements is paramount for restoring natural dexterity and allowing individuals with upper limb loss to perform a range of daily activities that require fine motor skills, including typing, operating tools, or grasping small objects. Moreover, existing research has utilized entire datasets while reporting overall accuracy yet failing to detail the sensitivity and specificity of each finger movement. These approaches overlook the possibility that some movements may be well recognized while others may not be. Therefore, this study focuses on the most complex movements, those of the fingers, to provide a comprehensive analysis of each movement. To address existing gaps in literature, we propose a tailored 1D-CNN approach for fine-resolution finger movement recognition. The distinction of our proposed method from existing deep learning approaches lies in an optimized balance between deep feature extraction and computational efficiency, achieved by pairing high-resolution signal preprocessing techniques with a streamlined 1D architecture. This will contribute to improving future recognition systems for prosthetic design and optimizing overall progress capabilities. Research has mainly focused on classifying individual finger movements, not whole hand gestures. For example, Tenore et al. [ 21 ] achieved over 98% accuracy in decoding 12 finger motions using surface EMG and neural networks with five participants. Kanitz et al. [ 22 ] used linear discriminant analysis and support vector machines on eight finger primitives with one participant, reaching 97.7% accuracy. Al-Timemy et al. [ 23 ] combined features with discriminant analysis to classify various finger movements in ten intact-limbed and six amputee subjects. Malesevic et al. [ 24 ] introduced a model for real-time finger movement decoding for prosthetics. These studies show that finger classification is possible but rely on simple features and small groups, and report mainly overall accuracy without detailed per-movement analysis. This study examines nine finger movements in a larger cohort of 40 subjects from the NinaPro DB2 dataset. It provides a systematic analysis of sensitivity, specificity, precision, and F1-score for each movement, addressing the clinical need for reliable finger movement decoding to restore dexterity and perform daily activities. Unlike prior work focused on overall accuracy, our approach uses a 1D-CNN with high-resolution preprocessing, offering detailed per-movement diagnostics. This method advances analytical depth, cohort size, and diagnostic granularity, aiming to improve prosthetic design and hand function recognition. 2. Methods The approach proposed in this paper is illustrated in Fig 1 . It begins with data as input, applies pre-processing techniques, and concludes with the classification of nine finger movements. Each corresponding subsection details the sequence of the methodology adopted. Download: PNG larger image TIFF original image Fig 1. Overview of the proposed methodology pipeline. https://doi.org/10.1371/journal.pone.0361050.g001 2.1 Data Acquisition Selecting an sEMG dataset is critical to ensure effective training and evaluation of hand gesture classifiers, as model accuracy and generalizability can vary widely from one dataset to another. The NinaPro DB2 dataset, comprising recordings from 40 participants performing 49 different hand movements using 12 surface electrodes, is a widely recognized benchmark in the prosthetic field [ 3 ]. The NinaPro DB2 dataset additionally provides inertial, kinematic, and force data to supplement comprehensive analysis. While other publicly accessible EMG datasets cover diverse demographics and amputee populations, NinaPro DB2 was determined to be the best fit. Among the range of movement types offered in DB2, this study focuses on D movements, which primarily involve finger movements. A total of nine finger movements from the available data were included, each displaying the EMG waveform and corresponding to a color code ( Fig 2 ). Download: PNG larger image TIFF original image Fig 2. Color-coded EMG signal waveforms corresponding to each of the nine finger movements. https://doi.org/10.1371/journal.pone.0361050.g002 2.2 Pre-Processing Several preprocessing steps have been taken to enhance sEMG signals. A band-pass filter (20–500 Hz) was applied to preserve EMG energy while attenuating unwanted frequencies, and a notch filter (50–60 Hz) was applied to suppress power-line interference [ 25 , 26 ]. The corresponding illustration, Fig 3 , shows the pre-processing steps in this paper. Download: PNG larger image TIFF original image Fig 3. Demonstration of the sEMG signal denoising process. https://doi.org/10.1371/journal.pone.0361050.g003 2.2.1 Denoising-based Wavelet. Wavelet-based denoising is a widely adopted method that decomposes EMG signals into multiple frequency components through wavelet transformations. Since signal noise typically involves high-frequency components, a threshold is applied: small coefficients are either reduced or set to zero, while significant signal coefficients are preserved [ 27 – 29 ]. The denoised and filtered signal is further reconstructed via an inverse wavelet transform. In this stage, we specified the mother wavelet (Daubechies 4) and the exact number of decomposition levels (5) applied before empirical Bayesian thresholding [ 28 , 30 ]. Fig. 4 presents the original and denoised sEMG signals after wavelet-based and Bayesian denoising. Download: PNG larger image TIFF original image Fig 4. Comparison between original and denoised sEMG signals after wavelet-based and Bayesian denoising techniques. https://doi.org/10.1371/journal.pone.0361050.g004 2.2.2 Rectification. Rectification is a commonly applied technique to EMG signals following band-pass filtering, where bipolar waveforms are converted into unipolar signals by extracting absolute values of the signal at designated time points. Full-wave rectification was applied, converting negative amplitudes to positive to ensure that all muscle activation energy is represented by positive magnitudes [ 31 ]. Fig 5 illustrates the rectified EMG signal. Download: PNG larger image TIFF original image Fig 5. Full-wave rectified sEMG signals. https://doi.org/10.1371/journal.pone.0361050.g005 2.2.3 Normalization. The normalization step is applied to maintain consistency in EMG signal amplitudes across channels, subjects, and recording sessions. In this paper, we strictly applied the Z-score standardization method (zero mean, unit variance) calculated independently for each of the 12 channels (per-channel normalization) to prevent higher-amplitude channels from dominating the feature space during 1D-CNN training by scaling signals relative to a reference contraction, enabling standardized interpretation of signal strength across subjects in a dataset [ 32 ]. Z-scores are used because they are less affected by isolated high-amplitude outliers. Fig 6 highlights the normalized signal. Download: PNG larger image TIFF original image Fig 6. Normalized sEMG signal after applying z-score standardization. https://doi.org/10.1371/journal.pone.0361050.g006 Where: μ is the mean of the training signal σ is the standard deviation of the training signal 2.3 Windowing A necessary step is implementing sliding-window segmentation to divide continuous EMG into overlapping short windows, aiming to achieve high hand gesture recognition accuracy. Each window serves as a sample for classification to capture temporal patterns of muscle activation, aiming to achieve accurate gesture recognition. Finding an optimal balance between information content and responsiveness is paramount: selecting a long window introduces potential phases or transitions, whereas a short window may fail to capture the full gesture pattern [ 33 ]. A 400 ms window with 50% overlap was selected to fully capture slower and more complex gestures. Despite a 400 ms window exceeding the typical 150–250 ms recommendation for a real-time myoelectric window, it was selected for this offline study because longer windows yield superior recognition for refined and coordinated finger movements with distributed EMG patterns. Small windows, on the other hand, tend to have difficulty distinguishing similar movements. This decision stemmed from qualitative rather than systematic evaluation, and no accuracy-latency trade-off analysis was performed. Exploring various window lengths (e.g., 150–400 ms) to optimize both performance and response time presents promising grounds for future real-time research [ 34 ]. The illustration below, Fig 7 , presents the optimized window selection. Download: PNG larger image TIFF original image Fig 7. The optimized window size with overlapping process. https://doi.org/10.1371/journal.pone.0361050.g007 2.4 Data Splitting To avoid temporal correlation, this study utilized repetitions 2 and 5 from the NinaPro DB2. The model was trained and tested on the same individual across different repetitions, commonly referred to as within-subject evaluation, as opposed to cross-subject evaluation, which uses data from multiple individuals. Recent literature suggests partitioning the data into about 60–70% for training, 10–20% for validation, and 10–30% for testing. Selecting repetitions 2 and 5 keeps assessment bias risk to a minimum by preventing the model from memorizing data patterns. 2.5 Convolutional Neural Network (CNN) Convolutional neural networks (CNNs) are widely used in EMG classification for their ability to extract deep features using optimal filters automatically. 1D-CNNs effectively process multichannel EMG signals over time by capturing both spatial and temporal features, including local motifs, thereby aiding hand gesture recognition. They are faster, simpler, and easier to train than RNNs such as LSTMs, offering a good balance of accuracy and efficiency, especially with refined regularization [ 35 ]. The 1D-CNN architecture starts with an input layer, followed by convolutional layers with filters and activation functions, a global pooling layer, a fully connected layer, a Softmax layer, and classification layers. It was trained with the Adam optimizer for 500 epochs and specific hyperparameters, demonstrating its effectiveness in EMG signal classification ( Fig 8 ). Download: PNG larger image TIFF original image Fig 8. Schematic of proposed 1D–CNN model architecture. https://doi.org/10.1371/journal.pone.0361050.g008 For hyperparameter tuning, we empirically optimized all parameters. We conducted an iterative evaluation of different configurations, selecting the combination that yielded the most stable convergence on the validation set without oscillations. To prevent overfitting, we employed a multi-layered approach: We applied L2 Regularization (weight decay) with a coefficient of 0.001 to penalize large weights. We enforced dynamic data shuffling (Shuffle: every epoch) to prevent the network from memorizing sequence patterns. We continuously monitored the validation loss through an independent validation set to ensure no generalization gap emerged between the training and testing phases. ϵ : input EMG segment with C = 12 channels, T = 800 time steps ϵ : weights of the convolution layer l : output feature map Where: f : filter index k : kernel index K : kernel size 3. Results The model is trained using the previous settings mentioned in the methods section; the results are summarized in the confusion matrix and the receiver operating characteristic (ROC) curve in Fig 10 . As shown in the confusion matrix in Fig 10a , the classifier’s overall accuracy across all movements was 90.1%. A total of 1,777 records were correctly classified, achieving a precision of 89.7% and a recall of 93.1% for the flexion of the little finger labeled “41”. Label “42” corresponds to the flexion of th
## Related Clinical Research

- [Correction: References updated in PLOS ONE meta-analysis of mind-body exercise for mild cognitive](https://medichelpline.com/clinical-feed/plos-one-0-correction-nonlinear-dose-response-of-mind-body-exercise-in-older-adults-with.md)
- [Otolin-1 lacks diagnostic discrimination for central versus peripheral acute continuous dizziness](https://medichelpline.com/clinical-feed/plos-one-9-otolin-1-for-differentiating-central-and-peripheral-causes-of-acute-continuous.md)
- [Performance of the CBI-M Framework in CENTER-TBI and TRACK-TBI Cohorts](https://medichelpline.com/clinical-feed/plos-medicine-0-the-cbi-m-framework-for-characterization-of-traumatic-brain-injury-evaluation.md)
- [FDA Launches NAIN to Accelerate Neuroscience Medical Product Development](https://medichelpline.com/clinical-feed/fda-news-releases-0-fda-launches-new-program-to-advance-neuroscience-medical-product-development.md)
- [APOLLAI AI versus EMS for prehospital detection of stroke deficits](https://medichelpline.com/clinical-feed/medrxiv-0-apollai-a-multimodal-ai-system-versus-ems-personnel-for-prehospital-detection.md)

## Navigation
- [← Back to Neurology Feed](https://medichelpline.com/clinical-feed/neurology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.