Aprendizagem de máquina aplicada a análise do movimento de membro superior de pessoas com esclerose lateral amiotrófica

Abstract

Amyotrophic Lateral Sclerosis (ALS) leads to gradually progressive motor limitations, which appear differently in each patient, such as impairments related to the upper limb (UL), in particular, strongly impacting the performance of activities of daily living and functional independence. Technological development has enabled the use of surface electromyography (EMGs) and the accelerometer (ACC) as additional tools for analyzing of motor function, thereby improving the existing assessment tools. These can still be enhanced with the association of machine learning (ML) algorithms that make them more accurate and precise for evaluation and can contribute to the improvement of assistive technology (AT) resources, such as orthoses. From this perspective, we aimed to implement and compare the ML algorithms on sEMG and ACC data to assess the UL motor function of people with ALS. The findings of this research were divided into four articles, as described below. The findings of this research were divided into four articles: article 1 consists of a scoping review whose objective was to describe the characteristics of UL orthoses controlled by ML algorithms, based on information extracted from articles and patents; and assess the risk of bias (PROBAST) of the articles. 16 articles and 4 patents were inserted. In this review, we identified that the ML-based orthoses have different physical and control characteristics with divergence in the usability tests performed. Linear regression models and sEMG are the most common control approaches and the risk of bias is classified as “unclear” and “high”. Articles 2 to 4 deal with the analysis of the UL movement, based on sEMG and ACC data collected during the execution of a cross-sectional study approved by the Ethics Committee of the UFRN Central Campus (CAAE: 25687819.3.0000.5537). Ten healthy people and 7 with ALS were evaluated using a standardized assessment form and validated assessment instruments to analyze the health condition of people with ALS and/or other neurological and/or musculoskeletal disorders. sEMG and ACC data were analyzed to identify the degree of stationarity, using mean, variance, and Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test; linearity, by calculating the standard deviation, Brock, Dechert & Scheinkman test (BDS) and nonlinear autoregressive exogenous test (NLARX); gaussianity from the skewness and kurtosis tests; and sparsity through the gini index. sEMG data were classified as non-linear (BDS test – significance level = 1 and detection ratio = 2.4; NLARX test – rejection decision = 1 and rejection ratio = 1.5), stationary (KPSS test - rejection decision = 0 and p-value = .1), non-normal (skewness test - s = -0.77; kurtosis test - h = 3.14) and sparse (g = 83). However, the ACC data were categorized as non-linear (BDS test – significance level = 1 and detection ratio = 29.21; NLARX test – rejection decision = 1 and detection ratio = 1.59), non-stationary (KPSS test - rejection decision = 1 and p-value = .01), non-normal (skewness test - s = -0.81; kurtosis test - h = 3.17) and sparse (g = 93). However, statistical differences were revealed only regarding the degree of linearity and stationarity of the sEMG and ACC data when comparing healthy people with ALS. When implementing classifiers to identify sEMG signals from both healthy and ALS people, it was identified that the support vector machine (SVM) yielded the best performance, based on the following observed metrics: sensitivity - 95.38%, accuracy - 96.74%, and UAR - 98.48%. The SVM trained with continuous wavelet transform (CWT) achieved the best metrics: accuracy - 97.35%, and UAR - 98.69%. ML-based UL orthoses may offer additional benefits on motor rehabilitation, which are movement changes caused by some diseases, such as ALS. People with ALS have statistical properties of the EMGs and ACC signal, level of fatigue, muscle activation, and range of movement that are different from those of healthy people. This information is essential for therapeutic planning and for the development of AT resources, which enable the improvement of functionality and independence.