Improving IgG determination in goat colostrum using colour through Artificial intelligence

Abstract

Introduction Immunoglobulin G colostrum estimation is crucial in ruminant dairy farms due to agammablogulinemic or hypogammablobulinemic status of the ruminant’s newborn. Some methods are use at farm, such as colostrometer or refractometer. Colostrum colour was proposed to estimate IgG in goat colostrum by Argüello et al. (2005). Since 2005, artificial intelligence has advanced exponentially. Algorithmic advancements, the exponential increases in computing power and storage and an explosion of data, as highlighted in the McKinsey AI guide (Chui & McCarthy, 2020), have evolved synergistically. This evolution has facilitated access to computational and statistical tools, opening a new realm of possibilities for the scientific community. The objective of the present study was improving the IgG estimation trough colour in goat colostrum using Machine learning and Deep learning methodologies. Materials and Methods The data set from Arguello et al. (2005) was used. IgG (using ELISA) and CIE colour (L, Cr and Hue) were measured in 813 goat colostrum samples. The data set was randomly divided in two subsets, one for training the models (650 registers) and one for testing the models (163 registers). The models have been generated using Python version 3.11.7. Techniques based on regression have been employed, specifically decision trees in Machine learning, utilizing the Sklearn package, and neural networks in Deep learning through the TensorFlow and Keras packages. Subsequently, in both techniques, the regression obtained has been factored into two values, HIGH in the case of IgG value greater than 20 mg/ml, or LOW otherwise, to generate evaluation metrics under the same conditions as those conducted in the original study. Results In the original 2005 study, an Accuracy of 0.87, a Sensitivity of 0.93, a Specificity of 0.71, and a Negative Predictive Value of 0.78 were obtained. Using the neural network-generated model, these values were to 0.96, 0.94, 0.97, and 0.97, respectively. Using tree regression, resulting in an Accuracy of 0.98, Sensitivity of 0.97, Specificity of 1, and Negative Predictive Value of 0.94. Conclusions The utilization of Machine Learning or Deep Learning techniques represents a significant improvement compared to classical statistical methods on the same sample data, obtaining regression data for IgG that are nearly equivalent to those obtained with laboratory techniques.