Electrophysiological Muscle Classification Using Multiple Instance Learning and Unsupervised Time and Spectral Domain...

Area of Science:

• Biomedical Engineering
• Neurology
• Machine Learning

Background:

• Electrophysiological muscle classification (EMC) is vital for diagnosing neuromuscular disorders.
• Current quantitative EMC methods lack clinical robustness and accuracy.
• There is a need for improved, unsupervised approaches to MUP analysis.

Purpose of the Study:

• To model EMC as a multiple instance learning (MIL) problem.
• To develop a system for unsupervised motor unit potential (MUP) labeling.
• To create a robust, supervised system for muscle classification.

Main Methods:

• Represent MUPs using morphological, stability, near fiber, and spectral features.
• Employ unsupervised Laplacian scores for MUP feature selection.
• Utilize neighborhood distance entropy consistency for MUP clustering (normal/abnormal).
• Represent muscles via MUP cluster associations.
• Classify muscles using support vector machines or random forests.

Main Results:

• The system successfully identified normal, myopathic, and neurogenic MUPs across four muscle types.
• Achieved an average classification accuracy of 98%.
• Outperformed previous EMC methods in accuracy.

Conclusions:

• Modeling EMC as MIL addresses the challenge of unsupervised MUP characterization.
• Incorporating morphological, stability, near fiber, and spectral features enhances muscle classification accuracy.
• The method effectively categorizes MUPs by disease without prior information.