Machine learning approaches to dietary classification from dental microtexture in primates

Area of Science:

• Paleontology
• Bioinformatics
• Computational Biology

Background:

• Dental microwear texture (DMT) analysis reconstructs mammal diets, but high-dimensional data complicates classification.
• Existing methods struggle with limited paleontological datasets and disparate parameter sets like ISO and scale-sensitive fractal analysis (SSFA).

Purpose of the Study:

• Develop a machine learning pipeline for automated classification of primate dietary groups and species using DMT data.
• Evaluate and compare the performance of various machine learning classifiers for dietary reconstruction.

Main Methods:

• Implemented a nested leave-one-out cross-validation framework to assess classifiers like multinomial logistic regression (MLR), Naive Bayes, and ensemble algorithms (Random Forests, XGBoost).
• Utilized ISO parameters, Fourier-based descriptors, and isotropy variables for feature extraction and model training.

Main Results:

• Lasso-regularized MLR and Naive Bayes achieved the highest predictive performance with strict feature selection for interpretability.
• Models using ISO parameters outperformed those using SSFA, indicating ISO variables better capture diet-specific micromechanical abrasions.
• Fourier-based descriptors and isotropy variables significantly improved model discrimination.

Conclusions:

• The developed machine learning pipeline offers a robust and reproducible method for accurate dietary classification from high-dimensional DMT data.
• This approach aids in resolving broader ecological questions, including niche partitioning, species evolution, and paleoecological dynamics.