Aprendizagem de máquina aplicada à execução da marcha em diabéticos tipo 2

Abstract

Introduction: Diabetes is characterized by a set of metabolic diseases that can cause several changes. One of them occurs in the sensorimotor function, which generates changes in gait execution, such as longer stance phase, shorter steps and inadequate plantar pressure distribution. Quantitative methods for assessing changes in gait patterns can be decisive in designing treatment strategies. Also, they can help in preventing complications caused by diabetes. With advances in machine learning (ML) techniques, automated pattern recognition in the face of massive amounts of data has become an essential tool in the medical field due to its ability to predict clinical complications before the disease gets worse. Objectives: To investigate ML models on gait assessment data from type 2 diabetic patients in order to identify gait patterns that may predict clinical complications of diabetes. Methods: The study involved two methodological phases: 1) Protocol and Systematic Review elaboration; 2) Development and improvement of predictive models of unsupervised and supervised BF for exploratory data analysis, detection of diabetes and detection of clinical complications in diabetes based on glycated hemoglobin (HbA1c) levels. The data for carrying out the study was provided through a partnership with Florida International University (FIU) during a sandwich doctorate (Edital No. 02/2020 – CAPES/PRINT) between September 2021 and June 2022. The data were pre-processed and implemented in different ML models. The ML models used were evaluated for their efficiency based on the silhouette analysis for unsupervised ML, metrics based on the confusion matrix for supervised ML, and conventional statistics, adopting a significance level of 5%. Results: Phase 1 resulted in two articles: Article 1 - The published protocol defined the methodology that guided the review; Article 2 - The systematic review, under consideration, resulted in four studies (208 participants) included. Two used ML as a predictive method, one used conventional statistics based on multiple stepwise regression, and one used the Fuzzy classifier, an uncertainty method. Studies achieved at least 75% in adequately reporting 19 Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) items. Three of the included studies were classified as high risk of bias. Phase 2 resulted in three articles in the submission process: Article 3 – K-Means separated the data set into two groups (silhouette = 0.47). Gait speed, step length and plantar pressure distribution patterns were statistically different (p < 0.05) between diabetics and non-diabetics. Among diabetics, there was a statistically significant difference (p < 0.05) in plantar pressure distribution patterns. Article 4 – Supervised ML algorithms using gait data showed high sensitivity in the distribution of plantar pressure in the heel region to classify diabetics from non-diabetics. Article 5 – The XGB classifier showed better results in classifying diabetes complications based on HbA1c levels, reaching an AUC of 0.99, a precision of 0.91, a recall of 0.90 and an f1-score of 0.89. For this classification, the most relevant gait characteristics were left support base, mean left pressure over time in the metatarsal region (I-III) and mean active sensor area in the phalanges (III-IV). Conclusion: The literature shows few studies using gait data as predictors of diabetes. Type 2 diabetics presented changes in gait performance, with differences in plantar pressure distribution in individuals with higher glycemic levels. Different regions of plantar pressure distribution were relevant in the classification of diabetics or non-diabetics and in detecting complications in diabetes. These findings have been supported in the literature.