Enhanced Detection of Hand Gestures From sEMG Signals Using Stacking Ensemble With Particle Swarm Optimization and Meta-Classifier

• Published in: IEEE access Vol. 13; pp. 63611 - 63626
• Main Authors: Chandra Sakinala, Umesh, Abinaya, S.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; CNN; Computational modeling; Convolutional neural networks; Data models; Datasets; Deep learning; Electromyography; Feature extraction; Gesture recognition; Hands; Human computer interaction; Long short term memory; LSTM; Machine learning; Muscles; Muscular function; Particle swarm optimization; random forests; Surface electromyography
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3559182

Hand gesture recognition is a key aspect of Human Computer Interaction (HCI), that enables recognition of hand movements or gestures through sensors and cameras. Surface Electromyography (sEMG) is a technique used to measure electrical signals generated by human muscle activity. Analyzing sEMG signals are highly influenced by the factors like electrode placement, muscle contraction patterns etc. Existing methodologies for detecting sEMG signals face challenges in interpreting signals and frequency feature variations, primarily due to factors such as noise and motion artifacts, which significantly limit the accuracy and reliability when applied to large sEMG datasets. To overcome these challenges, a cutting-edge deep learning framework is proposed, combining a convolutional autoencoder with a stacking ensemble method that integrates CNN and LSTM models, optimized through particle swarm optimization to fine-tune the hyperparameters for superior training performance. Further to reduce the prediction error, the model employs a Meta classifier technique which incorporates Random Forests approach for improved accuracy. It is evaluated on the NinaPro DB1 dataset with 27 gestures from 52 subjects and a window length of 500 ms. It achieves high accuracy of 85% and effectively mitigates the impact of erroneous data patterns for advancing the robustness of gesture recognition systems.