Evolutionary Denoising-Based Machine Learning for Detecting Knee Disorders

摘要

Surface electromyography (sEMG) is a non-invasive tool that can aid physiological assessment of knee disorders towards clinical interventions. Machine Learning (ML) is widely used to classify sEMG data to help with early detection of knee disorders; however, the inherent noise and the high non-linearity of sEMG signals make pattern recognition a challenging task. This study aims to partly overcome these challenges with existing ML-based classifiers by denoising sEMG signals further via an innovative two-fold evolutionary approach. A novel Genetic Algorithm-based denoising approach is applied to sEMG data to decrease the search space for pattern-related classification. Thereafter, the proposed denoising technique is coupled with an ML-based classifier to improve the discrimination between physiological and pathophysiological knee functions from sEMG data by optimising its hyperparameters too. Thus, the novel evolutionary approach serves two purposes. Firstly, it further reduces noise in sEMG signals via a new GA-based denoising technique to concurrently maximise mutual information and minimise entropy; secondly, it also enables the optimisation of the classifier’s hyperparameters. The classification performance of the resulting hybrid algorithm was validated using sEMG data on 144 subjects (67 patients with knee disorders, 77 healthy subjects) and was found higher (ACC = 99.57%, 95% CI: 99.47–99.66; AUC = 1, 95% CI: 0.98–1) than that of similar ML algorithms and published studies. The hybrid algorithm achieved the highest classification performance by leveraging an evolutionary approach for effective denoising and hyperparameter optimisation, whilst retaining the lowest computational cost. Thus, the proposed evolutionary denoising ML-based classifier is deemed an accurate and reliable decision support system to aid the detection of knee disorders.