Deep learning for waveform identification of resting needle electromyography signals

Area of Science:

• Medical data analysis
• Machine learning
• Artificial intelligence

Background:

• Advancements in AI and machine learning offer new tools for analyzing medical data.
• Electromyography (EMG) is a key diagnostic tool in neurology.

Purpose of the Study:

• To investigate the efficacy of deep learning algorithms in classifying resting needle electromyography (n-EMG) discharges.
• To determine if AI can aid in the interpretation of n-EMG signals.

Main Methods:

• Six clinical resting n-EMG signals were converted to Mel-spectrograms.
• Data augmentation techniques were applied to increase the training dataset size.
• Deep learning models were trained and evaluated with and without pre-trained weights (transfer learning).

Main Results:

• The deep learning model achieved an accuracy of 0.86 on the original dataset.
• Data augmentation, increasing the dataset to 200,000 images, resulted in a perfect accuracy of 1.0.
• Transfer learning (using pre-trained weights) outperformed training models from scratch.

Conclusions:

• Deep learning algorithms can effectively classify resting n-EMG signals.
• Data augmentation and transfer learning are crucial for maximizing classification accuracy.
• AI-powered tools show promise for computer-aided identification in clinical n-EMG testing.