eng Oriental Scientific Publishing Company Biomedical and Pharmacology Journal 0974-6242 2025-02-20 18 March Spl Edition 315 329 10.13005/bpj/3090 64117 article Thejaswini Kishore 1 Subin Sunderraj Remabai 1 Chitra Retnaswamy 1 Eben Sophia Paul 1 Division of Computer Science and Technology, Karunya Institute of Technology and Sciences, Coimbatore, India In order to solve important issues including inter-subject variability, noise interference, and electrode misalignment, this work presents a thorough hybrid deep learning architecture designed for electromyography (EMG)-based gesture detection. The suggested model sets a new standard in the field by achieving previously unheard-of levels of accuracy and robustness through the smooth integration of Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory networks (BiLSTM), and channel attention mechanisms. On benchmarks such as the NinaPro (Non-Invasive Adaptive Prosthetics) database, traditional methods for EMG-based gesture identification, which usually rely on hand-crafted features or stand-alone deep learning modules, reach high accuracies of roughly 93%. On the other hand, our architecture significantly outperforms current techniques by utilizing the distinct advantages of each of its constituent parts to achieve an accuracy of 96.45%. The CNN module captures complex local and global patterns across EMG channels, acting as a potent spatial feature extractor. In addition, by processing signal sequences in both directions, the BiLSTM module is excellent at temporal modelling, which allows the architecture to capture the dynamic relationships present in EMG data. By resolving inter-channel variability and reducing the effect of noisy or misaligned electrodes, the channel attention mechanism dynamically prioritizes the most pertinent EMG channels to further improve performance. Numerous tests on the NinaPro dataset demonstrate the suggested model's higher performance. Interestingly, it achieves great accuracy even when evaluated on unknown subjects, demonstrating remarkable generalization in cross-subject training contexts. Additionally, the architecture exhibits exceptional resistance to noise and electrode misalignment, two significant issues that frequently impede the implementation of EMG-based systems in practical settings. These developments highlight the model's adaptability and usefulness for applications in prosthetics, human-computer interaction, and rehabilitation systems. The suggested architecture offers a scalable and effective solution for real-time applications in addition to attaining improved performance metrics. Because of its versatility and lightweight construction, it can be used in areas with limited resources, including embedded systems or wearable technology. The model's ability to manage a variety of dynamic real-world situations is further strengthened by the incorporation of strong preprocessing and augmentation approaches.  By tackling the fundamental drawbacks of conventional and current deep learning techniques, this work establishes a new benchmark for EMG-based gesture identification. In addition to improving accuracy, the hybrid architecture's capacity to combine spatial, temporal, and attention-based elements into a coherent framework guarantees robustness and adaptability, opening the door for revolutionary applications in assistive technology, healthcare, and other fields. https://biomedpharmajournal.org/vol18marchspledition/enhanced-emg-based-gesture-recognition-using-hybrid-cnn-bilstm-architecture-with-channel-attention/ BiLSTM CNN Channel Attention Deep Learning EMG Gesture Recognition NinaPro Database