Investigators constructed a one-dimensional convolutional neural network (1D-CNN) specifically adapted to identify nine distinct finger movements from surface electromyography (sEMG) signals. Emphasis was placed on balancing deep feature extraction capabilities with computational efficiency—especially pertinent for future prosthetic applications constrained by low-power hardware. The network consisted of convolutional layers followed by activation, global pooling, fully connected, Softmax, and classification layers. Training utilized the Adam optimizer over 500 epochs, with L2 regularization and data shuffling to mitigate overfitting. Hyperparameters were empirically optimized to achieve stable convergence without oscillations, resulting in a model well-suited for multi-channel temporal EMG data processing.
Raw sEMG signals underwent multiple preprocessing steps to enhance signal quality and feature representation. Band-pass filtering (20–500 Hz) preserved relevant neuromuscular activity while attenuating extraneous frequencies. Notch filtering removed power-line interference (50–60 Hz). Wavelet-based denoising employed Daubechies 4 mother wavelet decomposition over five levels, followed by empirical Bayesian thresholding to suppress noise components effectively. Full-wave rectification converted bipolar waveforms to unipolar signals, ensuring all muscle activation magnitudes were positive. Signals were finally normalized per channel using z-score standardization to reduce amplitude variability across electrodes and subjects, thereby preventing dominant channel bias during model training.
Data were sourced from the established Ninapro DB2 database, which includes recordings from 40 participants performing a wide range of hand and finger movements with 12 surface electrodes. This study narrowed focus to nine finger movements (classifications labeled from 41 to 49), providing a refined scope relative to broad hand gestures. The choice enabled detailed assessment of individual finger motor control signals, which are critical for natural hand function restoration. The dataset's high-quality signal acquisition under standardized conditions supported robust modeling and comparative benchmarking.
Signal segmentation employed a 400 ms sliding window with 50% overlap, generating temporally coherent segments for input into the 1D-CNN. This window size was chosen to balance temporal resolution with the ability to capture complex, distributed EMG activation patterns associated with fine finger movements. Although longer than the typical 150–250 ms windows suggested for real-time control, this offline assessment prioritized classification accuracy for nuanced gestures over latency, supporting future refinements on response time versus accuracy trade-offs.
The model was evaluated within-subject, using repetitions 2 and 5 of the Ninapro DB2 dataset for training and testing, respectively. This approach avoided temporal correlation that can bias performance metrics while preserving individual muscle activation patterns essential for personalized prosthetic control. Data splits conformed to recommended partitioning ratios (60–70% training, 10–20% validation, 10–30% testing), minimizing overfitting potential and ensuring representative generalization on held-out repetitions.
Results indicated an overall classification accuracy of 90.1% across the nine finger movements. Precision and recall metrics were calculated for each class, revealing notably high classification fidelity for certain fingers, such as the little finger flexion with 89.7% precision and 93.1% recall, while other classes demonstrated varied specificity and sensitivity. Receiver operating characteristic (ROC) curves and confusion matrices illustrated classification performance nuances, informing on movements prone to misclassification. These granular diagnostics extend beyond prior studies that primarily reported global accuracy, offering deeper clinical insights essential for prosthetic control reliability.
Model design accounted for low computational complexity to accommodate deployment constraints typical of wearable prosthetic systems. The 1D-CNN architecture evidences a favorable trade-off between classification accuracy and resource demand, eschewing larger or more complex models like Vision Transformers, which require greater parameter counts and computational power beyond currently practical embedded systems.
Prior investigations have ranged from hand gesture classification to more limited finger movement recognition, often using simpler feature extraction or smaller cohorts. Reported accuracies mostly focused on overall percentages without in-depth per-movement analysis, sometimes relying on extensive subject-specific training. This study differentiates by leveraging a sizable cohort, comprehensive preprocessing, and a lightweight 1D-CNN to achieve fine-grained finger movement decoding with robust interpretability.
Authors noted the offline nature of the study and the use of longer window sizes may reduce applicability for real-time control, suggesting future research should explore latency-accuracy balances and window size optimization. Transformer-based models, despite promising accuracy for sEMG classification, currently pose computational challenges for embedded prosthetic hardware but remain an intriguing avenue for enhanced contextual modeling pending advancements in efficiency.
Overall, these findings provide foundational evidence for the viability of deep learning-enhanced, fine-resolution finger movement recognition from surface EMG signals, informing the development of next-generation upper-limb prostheses capable of nuanced motor function restoration.