Development of Artificial Intelligence to Support Needle Electromyography Diagnostic Analysis.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Diagnostic needle electromyography (EMG) generates resting membrane potential signals.
• Accurate classification of these signals is crucial for diagnosing neuromuscular disorders.
• Current methods may benefit from advanced image recognition techniques.

Purpose of the Study:

• To develop an artificial intelligence (AI)-based image recognition scheme for classifying EMG waveform images.
• To utilize TensorFlow-Slim and Python for implementing this AI scheme.
• To classify three types of abnormal resting membrane potential signals: positive sharp waves (PSWs), fibrillations (Fibs), and Others.

Main Methods:

• A convolutional neural network (CNN) model, Inception v4, was employed using the TensorFlow-Slim library.
• A dataset of 8,576 waveform images, derived from 4,015 raw EMG data instances, was used for training.
• The pretrained Inception v4 model was fine-tuned for image classification of EMG waveforms.

Main Results:

• The Inception v4 classification model achieved high performance: 93.8% accuracy, 99.5% precision, and 90.8% recall.
• The model successfully classified EMG waveform images into the three predefined categories.
• The image recognition model was trained using diverse medical image data.

Conclusions:

• TensorFlow-Slim enables straightforward training and recognition of medical image data, such as EMG waveforms, with simplified coding.
• This study demonstrates the potential of CNNs for analyzing electrophysiological signal waveforms represented as images.
• The developed AI scheme offers a promising tool for enhancing diagnostic EMG interpretation.