TY - GEN
T1 - Muscle categorization using PDF estimation and Naive Bayes classification
AU - Adel, Tameem M.
AU - Smith, Benn E.
AU - Stashuk, Daniel W.
PY - 2012
Y1 - 2012
N2 - The structure of motor unit potentials (MUPs) and their times of occurrence provide information about the motor units (MUs) that created them. As such, electromyographic (EMG) data can be used to categorize muscles as normal or suffering from a neuromuscular disease. Using pattern discovery (PD) allows clinicians to understand the rationale underlying a certain muscle characterization; i.e. it is transparent. Discretization is required in PD, which leads to some loss in accuracy. In this work, characterization techniques that are based on estimating probability density functions (PDFs) for each muscle category are implemented. Characterization probabilities of each motor unit potential train (MUPT) are obtained from these PDFs and then Bayes rule is used to aggregate the MUPT characterization probabilities to calculate muscle level probabilities. Even though this technique is not as transparent as PD, its accuracy is higher than the discrete PD. Ultimately, the goal is to use a technique that is based on both PDFs and PD and make it as transparent and as efficient as possible, but first it was necessary to thoroughly assess how accurate a fully continuous approach can be. Using Gaussian PDF estimation achieved improvements in muscle categorization accuracy over PD and further improvements resulted from using feature value histograms to choose more representative PDFs; for instance, using log-normal distribution to represent skewed histograms.
AB - The structure of motor unit potentials (MUPs) and their times of occurrence provide information about the motor units (MUs) that created them. As such, electromyographic (EMG) data can be used to categorize muscles as normal or suffering from a neuromuscular disease. Using pattern discovery (PD) allows clinicians to understand the rationale underlying a certain muscle characterization; i.e. it is transparent. Discretization is required in PD, which leads to some loss in accuracy. In this work, characterization techniques that are based on estimating probability density functions (PDFs) for each muscle category are implemented. Characterization probabilities of each motor unit potential train (MUPT) are obtained from these PDFs and then Bayes rule is used to aggregate the MUPT characterization probabilities to calculate muscle level probabilities. Even though this technique is not as transparent as PD, its accuracy is higher than the discrete PD. Ultimately, the goal is to use a technique that is based on both PDFs and PD and make it as transparent and as efficient as possible, but first it was necessary to thoroughly assess how accurate a fully continuous approach can be. Using Gaussian PDF estimation achieved improvements in muscle categorization accuracy over PD and further improvements resulted from using feature value histograms to choose more representative PDFs; for instance, using log-normal distribution to represent skewed histograms.
UR - http://www.scopus.com/inward/record.url?scp=84882938907&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84882938907&partnerID=8YFLogxK
U2 - 10.1109/EMBC.2012.6346501
DO - 10.1109/EMBC.2012.6346501
M3 - Conference contribution
C2 - 23366462
AN - SCOPUS:84882938907
SN - 9781424441198
T3 - Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS
SP - 2619
EP - 2622
BT - 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBC 2012
T2 - 34th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS 2012
Y2 - 28 August 2012 through 1 September 2012
ER -