USO DE TÉCNICAS DE APRENDIZADO DE MÁQUINA PARA ESTIMAÇÃO DE GORDURA CORPORAL EM ADULTOS BRASILEIROS

Abstract

Considering the rising prevalence of obesity in Brazil and worldwide, the limitations of Body Mass Index (BMI) in differentiating lean mass from fat mass, as well as the absence of a precise and accessible method to assess body fat percentage (%BF), this study aimed to develop models capable of estimating %BF in adults, using demographic data and simple anthropometric measures, through Machine Learning (ML) techniques. All statistical analyses and model construction were performed using the R® programming language. The study population consisted of 7,085 adults aged 22 and 30 years, belonging to the 1993 and 1982 Pelotas-RS birth cohorts, respectively. Participants' BMI values were calculated and classified according to the World Health Organization. Individuals were further categorized based on %BF values into obese (≥ 25% for men and ≥ 32% for women) and non-obese (≤ 25% for men and ≤ 32% for women). After BMI and %BF classification, individuals with BMI < 25 kg/m² were also grouped into normal weight (BMI < 25 kg/m² and absence of obesity by %BF) and normalweight obese (BMI < 25 kg/m² and presence of obesity by %BF). Input variables considered were demographic data (sex and age) and anthropometric measures (weight, height, and waist (WC), hip (HC), right wrist, and right calf circumferences). The outcome was defined as the %BF measured by Dual-Energy X-ray Absorptiometry (DXA). For model construction, the Radial Support Vector Machine (SVM) algorithm was applied, adopting Linear Regression (LR) as a baseline. Data partitioning followed the Hold-Out method, with 80% comprising the training set and 20% the test set. Exclusively in the training group, the k-fold Cross-Validation technique was applied, adopting k = 5. To evaluate performance, the following metrics were calculated: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Normalized Root Mean Square Error (NRMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²). Subsequently, in the best-performing model, the SHAP function was applied to select the most relevant variables for constructing two reduced versions: Reduced 1 (containing only the five most relevant variables based on SHAP calculation – sex, WC, HC, height, and weight, in order of importance) and Reduced 2 (including the variable age). The Bland-Altman plot and Intraclass Correlation Coefficient (ICC) were applied to evaluate the agreement between the %BF estimated by the models and the %BF measured by DXA. The frequency of overweight was 26.5% and obesity 14.1%, considering BMI. However, by %BF, 52.1% of individuals were classified as obese. Notably, normal-weight obese individuals corresponded to 17.3% of the general sample. All ML models yielded better results compared to LR, with the Radial SVM - Reduced 2 achieving the best performance and agreement (MAE = 2.98; RMSE = 3.72; NRMSE = 7.07; R² = 0.91, MAPE = 13.24% and ICC = 0.95 [95% CI: 0.947 – 0.957]). It is concluded that the developed ML models, using simple demographic and anthropometric variables, presented high performance and excellent agreement for estimating %BF in adults, being potentially applicable in epidemiological research and clinical settings, including those with scarce resources, especially in Primary Health Care. It is noteworthy that the Radial SVM – Reduced 2 model proved to be accurate and parsimonious, requiring a smaller number of variables for its application.