Estimation of total body fat using symbolic regression and evolutionary algorithms

A 2025 EvoApplications paper compares evolutionary methods that search for readable formulas estimating total body-fat percentage. The methods are tested against the QLattice framework.

Published by Springer in April 2025 in the proceedings of the International Conference on the Applications of Evolutionary Computation, part of EvoStar, which met in Trieste, Italy, the paper by José-Manuel Muñoz, Odin Morón-García, Omar Costilla-Reyes, and J. Ignacio Hidalgo applies symbolic regression to the estimation of total body-fat percentage. The authors describe body-fat percentage as a more informative measure than body-mass index, which cannot distinguish fat mass from lean mass. Symbolic regression searches for a mathematical expression that fits data, producing a formula that can be read and checked rather than a set of model weights.

Three grammar-based evolutionary methods, Grammatical Evolution, Context-Free Grammar Genetic Programming, and Dynamic Structured Grammatical Evolution, are evaluated on a public anthropometric dataset, with the QLattice framework used as a comparison. Evolutionary methods improve a population of candidate formulas through repeated selection and variation, and the grammar restricts which formulas can be built so that every candidate stays valid and readable.

The experiments examine the trade-off between predictive performance and readable equations: more complex expressions may fit better but become harder for a clinician to interpret. The authors report that the grammar-based methods were competitive in both accuracy and interpretability. The reported models estimate body-fat percentage from measured body characteristics in the selected dataset.

The 2026 preprint on using large language models to audit the physiological plausibility of symbolic-regression formulas extends this study with several of the same collaborators. The paper also sits alongside Dr. Costilla Reyes’s symbolic-regression research at MIT, including the learned concept library (LaSR) and InceptionSR.