Development of a Model using Machine Learning Intended to be Embedded in a Wearable Device to Detect Muscle Fatigue based on sEMG Data Associated with a Sustained Single 80% Maximum Voluntary Contraction

• Author / Creator

  Islam, Md Manirul

• Background: Muscle fatigue is the progressive reduction in a muscle's ability to contract and exert force when performing a sustained task. Muscle fatigue may prevent the task from being complete and increase the risk of injury. Eventually, the performance of individuals during athletic activities can be limited due to fatigue. Individuals typically have to rely on their own perception of muscle fatigue and often report this to their trainer subjectively. Fatigue onset may occur differently from day to day due to the factors such as dehydration and electrolyte balance. As a result, constant vigilance is required to optimize exertion levels completing a task. Obtaining a balance between enhancing performance and preventing injury is essential in planning the desired exercise and training program for a specific activity. A wearable device to detect muscle fatigue in real-time can help define training strategies for optimal test-specific activities and exercises.
  Objective: The overall objective of this thesis was to develop a model using machine learning that can be embedded in a wearable device to detect muscle fatigue in real-time based on the sEMG data associated with a sustained single 80% maximum voluntary contraction (MVC).
  Five specific objectives were undertaken sequentially to accomplish the overall objective. The individual specific objectives included: 1) To conduct a scoping review to identify all machine learning algorithms that are potentially capable of detecting muscle fatigue in a real-time and computationally adaptable to be embedded in a wearable device, 2) To extract features from the previously recorded sEMG data, and select the most promising features that are associated with a pattern of fatigue, 3) To evaluate the performance of the algorithms selected in Objective 1 using the extracted features identified in Objective 2, 4) To then select the most promising algorithm based its accuracy in classifying fatigue state, and 5) To develop a model using the identified ML algorithm that has the potential to be embedded in a wearable device.
  Methods: To address our first objective, a scoping review involving six different electronic databases was undertaken, where a total of 67 studies were included.
  To address objectives 2-5, an sEMG dataset from 100 participants who performed a sustained single 80% MVC was evaluated. The Fast Fourier Transform was used to estimate the power spectrum. Several frequency and time domain features (RMS, IEMG, Power, Median and Mean Frequency) from the sEMG signal were extracted. The Neighbourhood Component Analysis (NCA) was applied to select promising features. After training each algorithm utilizing selected features, their performance in fatigue classification was evaluated. The most promising algorithm was selected to develop our proposed model based on the classification performance and adaptability to integrate the algorithm in a wearable device.
  Finally, a new model was developed, and its potential to be embedded in a wearable device detecting muscle fatigue in real-time was assessed.
  Results: The scoping review suggested four potential algorithms (LDA, LR, SVMs, and Ensemble) that had the potential to be integrated into a wearable device with the goal of fatigue detection in real-time.
  From the extracted sEMG features, NCA selected 14 features that were used to train the ML algorithm. Comparing the performance of each algorithm with selected features, the proposed OSVM model achieved the highest classification accuracy of 99.2% with a sensitivity of 99% and specificity of 99.2%. The area under the curve (AUC) for both fatigue and non-fatigue conditions achieved a maximum unity of 1.00. Further testing of the trained model with new data demonstrated a 100% fatigue classification accuracy, an outstanding performance in the current literature. Additionally, our developed OSVM model showed a 92% fatigue classification accuracy when using only the 5 most prominent features.
  Conclusion: This research represents a comprehensive automated method where the developed model can be used in the laboratory setting as well as a wearable device to detect muscle fatigue in real-time during a sustained single isometric task. Furthermore, the features reduction mechanism facilitates the model to perform adaptively on the criteria of fatigue forecasting time and performance accuracy. Based on the technical requirements, the model is suitable for embedding into the ‘Raspberry Pi Zero-W’ microcontroller. Deploying into the microcontroller, it is possible to be used as a wearable device to detect muscle in real-time.

• Subjects / Keywords

  • Surface EMG
  • MVC
  • Muscle Fatigue
  • Algorithm
  • Feature extraction
  • Feature selection
  • Model
  • Wearable device
  • Real-time

• Graduation date

  Fall 2021

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-1xn3-7g92

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.