Attention-Enhanced Deep Ensemble Network for Robust EMG Classification: A Comparative Study Against Traditional Machine Learning Pipelines

Clinically, telling the difference between myopathy
 and neuropathy is crucial because their symptoms are similar
 but come from different diseases. This research introduces a
 hybrid deep learning framework that combines Convolutional
 Neural Networks (CNN), Bidirectional Long Short-Term Mem
ory (BiLSTM), and Multi-Head Attention to efficiently capture
 spatial and temporal relationships in electromyography (EMG)
 inputs. We use handcrafted statistical features and frequency
domain attributes via Fast Fourier Transform (FFT) to make
 the model more stable. We also use focal loss to fix class
 imbalance and advanced data enhancement methods like time
 warping, magnitude warping, and Gaussian noise. The model
 undergoes evaluation by hold-out and 5-fold stratified cross
validation employing ensemble voting. The results of experiments
 using EMG data from the Deltoid muscle in healthy, neurogenic,
 and myopathic classes show that the method works better than
 others. It achieved 99.00% accuracy and a macro F1-score of
 0.99 in the hold-out setting and 98.70% accuracy with consistent
 F1-scores in cross-validation. These findings exceed traditional
 systems like Random Forest and MLP improved by Bayesian
 methodologies. Our research shows that this is the first ensemble
based attention-driven approach for EMG classification that uses
 both time- and frequency-domain features with focal loss. The
 proposed method has a lot of effectiveness to be used in real-time
 diagnostic and wearable medical systems.