Uso de modelos estatísticos e técnicas de aprendizado de máquina para predição do percentual de gordura

• Autor(a) principal: SILVA , Michele Bezerra
• Data de Publicação: 2023
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFMA
• Texto Completo: https://tedebc.ufma.br/jspui/handle/tede/tede/5341
• topic: gordura corporal;; antropometria;; modelos preditivos;; body fat;; anthropometry;; predictive models.; Saúde Pública

ABSTRACT Obesity, characterized as excess body fat, is an epidemic problem, with impacts on the quality of life of individuals and an increase in public health expenditure. Among the goals for combating obesity proposed by the World Health Organization is an emphasis on prevention and diagnosis in primary care. The aim of this thesis was to develop prediction models to estimate the percentage of body fat (%BF) in adolescents and young adults, based on simple anthropometric measurements. The results obtained enabled two articles to be written: "Predicting the percentage of fat in adolescents from anthropometric measurements using artificial neural networks" and "Assessing the accuracy of predictive equations based on simple anthropometric measurements in the diagnosis of obesity". The aim of the first article was to develop Artificial Neural Network (ANN)-based models to estimate body fat percentage (BFP) in adolescents using simple anthropometric measurements. This was a cross-sectional study with 2,155 adolescents (18 and 19 years old). Fat percentage was measured by air displacement plethysmography (ADP). Statistical analyses were performed in the R program, version 4.3.0. Different generalized and sex-specific feedforward ANN models were implemented with different combinations of sex, age (years), weight (kg), height (cm), waist circumference (WC), body mass index (BMI) and waist-to-height ratio (WHtR) variables. The generalized ANN models showed better performance (R²>0.75) compared to the models by sex (R² ≤0.72). WC was a variable of high importance, especially among boys. No differences were observed between measured and estimated fat percentage by artificial neural network (p>0.05). The ANN models developed from simple anthropometric measurements were effective in predicting %FG in adolescents. The incorporation of WC showed different responses in the estimation performance of %BF for boys and girls. The second article aimed to evaluate the accuracy of predictive equations for the diagnosis of obesity. Cross-sectional study with 3,103 individuals (18 to 23 years old) of both sexes. The reference method for measuring %BF and statistical program were the same as in the first article. Multiple linear regression (MRL) was developed to elaborate the equations, using age, sex, in addition to the simple anthropometric variables, xiv described in the first article. The developed equations showed a good predictive performance (R2≈0.80 and NRMSE≈0.09). In addition, they showed greater predictive ability for obesity diagnosis (specificity ≈ 0.64; false positive= 0.36; AUC≈0.9) when compared to BMI alone (specificity ≈ 0.39; false positive= 0.89 ; AUC≈0.2 ). The equations developed from measures that are easily applicable to clinical practice showed a high predictive ability and a greater ability to identify people with obesity than BMI. The equations developed in this study and may be useful for estimating %BF in the absence of the reference method and can be used as a practical tool for monitoring %BF and diagnosing obesity.