# Comparative application of machine learning approaches for body weight prediction in non-descript indigenous goats at different growth stages

## TL;DR

This study compares machine learning models to predict body weight in goats at different growth stages, showing better accuracy than traditional methods.

## Contribution

Demonstrates the superior performance of CART over MARS and stepwise regression for body weight prediction in indigenous goats.

## Key findings

- CART outperformed other models with R² values of 0.87, 0.94, and 0.99 at birth, weaning, and yearling stages.
- Stepwise regression had lower accuracy and higher error rates compared to machine learning approaches.
- Heart girth and body length showed strong correlations with body weight across all growth stages.

## Abstract

Accurate prediction of body weight (BW) in goats is vital for breeding, feeding, drug administration, and marketing decisions, particularly in resource-limited farming systems where weighing scales are often unavailable. Traditional regression models have been applied but are limited by multicollinearity and non-linearity in body measurement data. This study aimed to evaluate the predictive performance of two machine learning (ML) approaches – Classification and Regression Trees (CART) and Multivariate Adaptive Regression Splines (MARS) – for estimating BW in non-descript indigenous goats across birth, weaning, and yearling stages, compared with stepwise regression models.

A total of 100 goats were assessed at three growth stages: Birth (24 h), weaning (4 months), and yearling (12 months). Linear body measurements, body length (BL), sternum height, heart girth (HG), rump height, and withers height, were recorded alongside BW. Correlation analyses, stepwise regression, CART, and MARS models were developed. Model performance was evaluated using the coefficient of determination (R2), Pearson’s correlation coefficient (r), Akaike information criterion (AIC), and relative root mean square error (RMSE).

BW showed strong positive correlations with HG and BL across all stages, while associations varied with other morphometric traits. Stepwise regression models exhibited lower predictive power, as indicated by reduced R² values and higher RMSE and AIC scores. In contrast, ML approaches demonstrated superior accuracy. CART consistently outperformed MARS, with R2 values of 0.87, 0.94, and 0.99 at birth, weaning, and yearling, respectively. CART also exhibited the highest r values (up to 0.99) and lowest RMSE across training and test datasets.

ML techniques, particularly CART, provide robust and reliable prediction of BW in non-descript indigenous goats, surpassing conventional regression methods. These approaches can guide practical herd management decisions, including optimized feed allocation, drug dosage, and breeding selection, especially in resource-limited settings. While the study underscores CART’s effectiveness, further validation with larger datasets and additional morphometric traits is recommended to enhance generalizability.