Rule Discovery in Milk Content towards Mastitis Diagnosis: Dealing with Farm Heterogeneity over Multiple Years through Classification Based on Associations.

Abstract

Simple SummaryInvisible (subclinical) mastitis decreases milk quality and production. Invisible mastitis is linked to an increased use of antimicrobials. The risk of the emergence of antimicrobial-resistant bacteria is a major public health concern worldwide. Therefore, early detection of infected cows is of great importance. Machine learning has opened a new avenue for early mastitis prediction based on simple and accessible milking parameters, such as milk volume, fat, protein, lactose, electrical conductivity (EC), milking time, and milking peak flow. However, farm heterogeneity is a major challenge where multiple patterns can predict mastitis. Here, we employed a classification based on associations and scaling approach for multiple pattern discovery over multiple years. The approach we have developed helps to address farm heterogeneity and generalise machine learning-based diagnosis of mastitis worldwide.Subclinical mastitis, an economically challenging disease of dairy cattle, is associated with an increased use of antimicrobials which reduces milk quantity and quality. It is more common than clinical mastitis and far more difficult to detect. Recently, much attention has been paid to the development of machine-learning expert systems for early detection of subclinical mastitis from milking features. However, differences between animals within a farm as well as between farms, particularly across multiple years, are major obstacles to the generalisation of machine learning models. Here, for the first time, we integrated scaling by quartiling with classification based on associations in a multi-year study to deal with farm heterogeneity by discovery of multiple patterns towards mastitis. The data were obtained from one farm comprising Holstein Friesian cows in Ongaonga, New Zealand, using an electronic automated monitoring system. The data collection was repeated annually over 3 consecutive years. Some discovered rules, such as when the milking peak flow is low, electrical conductivity (EC) of milk is low, milk lactose is low, milk fat is high, and milk volume is low, the cow has subclinical mastitis, reached high confidence (>70%) in multiple years. On averages, over 3 years, low level of milk lactose and high value of milk EC were part of 93% and 83.8% of all subclinical mastitis detecting rules, offering a reproducible pattern of subclinical mastitis detection. The scaled year-independent combinational rules provide an easy-to-apply and cost-effective machine-learning expert system for early detection of hidden mastitis using milking parameters.