Electromyogram (EMG) signals can provide valuable insights into muscle activation and health;​ however, achieving high-quality signals is challenging. EMG signals are susceptible to contamination, and the acquisition of high-quality EMG signals often requires lengthy setup times and experienced technicians. EMG signal quality analysis is an important domain of research to increase EMG integration into clinical practice, wearable devices, and long-term monitoring. This thesis develops techniques to assess EMG signal quality to improve the ease of EMG acquisition.

A literature review of methods for EMG signal quality analysis was conducted. Key areas for further research were identified, including improved selection of high-quality channels within EMG arrays.

A process for establishing a ground truth of EMG signal quality based on expert human ratings was developed and used to label an EMG dataset. The intra-rater reliability ranged from 0.76–0.90, and the inter-rater agreement was 0.83, indicating good reliability and that the human rated datasets could be used for future assessments.

An automated method of quantifying EMG signal quality was developed in two stages. In Stage I, poor-quality EMG channels were detected and reconstructed. An interpolation-based method of detecting and reconstructing poor-quality channels in high-density EMG (HD-EMG) was developed. Compared with two other detection methods (root mean square and normalized mutual information), the interpolation-based method was most effective for identifying poor channels in HD-EMG (precision 95.0%, recall 100%). Reconstruction of poor-quality channels was best performed with a 2-dimensional spline interpolation (correlation = 0.99).

In Stage II, machine learning was explored to distinguish between adequate, good, and excellent channels. A random forest regression model provided the best overall result with Spearman correlations of 0.90, 0.91, 0.74, and 0.73 for inter-electrode distances of 10, 20, 30, and 40 mm, respectively. The recommended EMG feature set was an amplitude-based feature (e.g., root mean square) combined with frequency-based feature(s) (e.g., mean frequency). In single channel EMG, the Spearman correlation dropped to 0.55;​ however, there was potential to distinguish between adequate and excellent quality channels.

The automated methods for quantifying EMG signal quality developed in this thesis will lead to improved efficiencies and lower costs for EMG acquisition.