Evaluation of machine learning models for predicting mastitis in dairy cows using dielectric properties and other milk quality parameters.

Early and accurate detection of mastitis is essential for improving udder health, milk quality and dairy farm productivity. This study evaluated machine learning models for estimating somatic cell count (SCC) from physico-chemical milk parameters integrated with dielectric properties and for assigning SCC derived udder health categories. A total of 982 longitudinal milk samples obtained from 100 dairy cows at LUVAS cattle farm (HISAR, HARYANA) were analyzed. Model inputs comprised pH, electrical resistance, fat, protein, lactose, milk temperature, udder surface temperature, and dielectric properties measured at six microwave frequencies (54, 104, 254, 500, 915, and 2450 MHz). Experimentally measured SCC served as target variable and was not included among the input predictors. After prediction, measured and predicted SCC were grouped using thresholds of <2 × 105, 2 × 105-5 × 105 and > 5 × 105 cells/ml. Feed Forward Neural Network (FFNN), Support Vector Machine (SVM) and Multiple Linear Regression (MLR) models were developed and evaluated using mean square error (MSE), correlation coefficient (R), sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). SVM demonstrated highest predictive performance (MSE = 2152.196, R = 0.98), followed by FFNN (MSE = 2447.60, R = 0.97), while MLR model exhibited lower accuracy (MSE = 13,231.630, R = 0.89). For SCC-based classification, FFNN achieved 100% sensitivity, 97.2% specificity and 0.986 AUC. SVM achieved 100% sensitivity, 98.7% specificity and 0.994 AUC, whereas MLR achieved 85.7% sensitivity, 98.6% specificity and 0.922 AUC. As repeated samples came from same cows, all samples were collected from one farm without external validation. These findings should be considered preliminary.