ABSTRACT: The aim of this work is to study the influence that genetics and physical condition have or may have in sports, specifically in cross training, with target variables related to power and percentage of fat-free mass. We work with data generated in Wingate tests, with a duration of 30 seconds; these are physical exercises at the maximum intensity possible in an ergometer. During the performance of this test, physical predictors that give an idea of the individual’s anaerobic capacity are measured. Two tests of this type were carried out, with a 5 minute interval during which both the above-mentioned physical predictors and the target variables of interest were continuously monitored. Prior to the test, the genetic variants (SNPs) presented by the individuals were analyzed for a set of 209 genes, along with physical predictors of anthropometric character. To achieve this objective, different machine learning models were studied: Random Forest, Bayesian Networks, traditional and optimized Multiple Linear Regression (with Best Subset Selection, Principal Components Analysis and Partial Least Squares) and K-Nearest Neighbors. It was found that the truly interesting relationships appear when both genetic and physical predictors are combined.
RESUMEN: El objetivo del presente trabajo consiste en estudiar la influencia que tienen o pueden tener la genética y la condición física en el ámbito deportivo, concretamente en cross training, con variables objetivo relacionadas con la potencia y el porcentaje de masa libre de grasa. Se trabaja con datos generados en pruebas Wingate, con una duración de 30 segundos; son ejercicios físicos a la máxima intensidad posible en un ergómetro. Durante la realización de esta prueba se miden predictores físicos que dan una idea de la capacidad anaeróbica del individuo. Se realizaron dos pruebas de este tipo, con un intervalo de 5 min durante las mismas, haciendo un seguimiento continuo tanto de los predictores físicos anteriormente mencionados como de las variables objetivo de interés. Previamente a la prueba se analizaron las variantes genéticas (SNPs) que presentaban los individuos para un conjunto de 209 genes, junto con predictores físicos de carácter antropométrico. Para la consecución de este objetivo se estudiaron diferentes modelos de machine learning: Random Forest, Bayesian Networks, Regresión Lineal Multiple tradicional y optimizada (con Best Subset Selection, Principal Components Analysis y Partial Least Squares) y K-Nearest Neighbors. Se encontró que las relaciones verdaderamente interesantes aparecen cuando se combinan tanto los predictores genéticos como físicos.
