Applications of Machine Learning in Identification and Prediction of Muscle Atrophy using EMG signals
Session Number
MEDH 13
Advisor(s)
Ashwin Mohan, Illinois Mathematics and Science Academy
Discipline
Medical and Health Sciences
Start Date
17-4-2025 2:45 PM
End Date
17-4-2025 3:00 PM
Abstract
Machine learning (ML) has significantly enhanced EMG signal processing, which has eased neuromuscular diagnosis and movement analysis. Previous studies focused on developing a variety of different ML models like Bayesian Techniques, Artificial Neural Networks, and Recurrent Neural Networks without a standardized way of comparing. Our work aims to contrast the performance between supervised, unsupervised, and neural network models in a standardized environment to accurately compare their effectiveness. Musculoskeletal and surface-EMG (sEMG) measurements were collected for four daily activities that measured gesture, movement, posture, and activity recognition. The signals were analyzed to estimatevarious parameters like FFT, Average, and RMS. The data along with participant demographic information such as age, gender, and self-reported physical activity level were used to train the model. The processed data served as the input to a Random Forest Classifier with 100 trees, a Support Vector Model with a linear kernel, and an Artificial Neural Network Model with a 64- neuron input unit, 32-neuron hidden unit, and 1 output, all using the sklearn library, and tested using the remaining 20% of the data. The Random Forest Classifier predicted the progression of the disease 3.9% more effectively than the ANN and 9.4% more effectively than the SVM.
Applications of Machine Learning in Identification and Prediction of Muscle Atrophy using EMG signals
Machine learning (ML) has significantly enhanced EMG signal processing, which has eased neuromuscular diagnosis and movement analysis. Previous studies focused on developing a variety of different ML models like Bayesian Techniques, Artificial Neural Networks, and Recurrent Neural Networks without a standardized way of comparing. Our work aims to contrast the performance between supervised, unsupervised, and neural network models in a standardized environment to accurately compare their effectiveness. Musculoskeletal and surface-EMG (sEMG) measurements were collected for four daily activities that measured gesture, movement, posture, and activity recognition. The signals were analyzed to estimatevarious parameters like FFT, Average, and RMS. The data along with participant demographic information such as age, gender, and self-reported physical activity level were used to train the model. The processed data served as the input to a Random Forest Classifier with 100 trees, a Support Vector Model with a linear kernel, and an Artificial Neural Network Model with a 64- neuron input unit, 32-neuron hidden unit, and 1 output, all using the sklearn library, and tested using the remaining 20% of the data. The Random Forest Classifier predicted the progression of the disease 3.9% more effectively than the ANN and 9.4% more effectively than the SVM.