Machine Learning with Applications (Jun 2022)
Deriving mapping functions to tie anthropometric measurements to body mass index via interpretable machine learning
Abstract
This paper aims at leveraging recent advancements in interpretable machine learning to better understand how anthropometric measurements can be tied to body mass index (BMI). Two objectives are of interest to this work, the first is to develop a properly validated interpretable machine learning (ML) ensemble capable of accurately predicting BMI, and the second is to derive a mapping function (i.e., a ML-based expression) that can describe the relationship between BMI and anthropometric measurements. This paper analyzes a historical database published by Penrose et al. (1985) containing thirteen body circumference measurements for 252 men. Four ML algorithms are then blended into an ensemble to examine the collected anthropometric measurements, namely, Extreme Gradient Boosted Trees, Light Gradient Boosted Trees, Random Forest, and Keras Slim Residual Network. The ensemble was then augmented with the SHAP (SHapley Additive exPlanations) and partial dependence plot techniques to understand the effect of each anthropometric measurement on BMI and the interaction between these anthropometric measurements. The proposed ensemble not only can comprehend the relation between anthropometric measurements and BMI index and derive a new non-parametric expression to predict BMI, but can also be used to interpret such relation and help us understand the logic driving ML’s predictions. Further, the interpretability analysis reveals that the main anthropometric measurements influencing BMI are the chest, abdomen, and hip circumference. Finally, clinicians and researchers are encouraged to leverage interpretability ML tools in their works instead of those of “black-box” nature.
