Developing machine learning models for fluid milk spoilage classification.

The dairy as well as other food industries face major challenges as retirement of experienced staff members can lead to substantial knowledge and skill gaps. New artificial intelligence and modeling tools may offer an opportunity for the dairy industry to capture existing expert knowledge into digital systems. A specific industry challenge includes the identification of fluid milk spoilage patterns, which is a pre-requisite for selecting and implementing targeted spoilage control strategies; traditionally, these classification tasks have been performed by human experts. To address this challenge, we developed a machine-learning-based digital expert system that uses microbiological data to classify the major types of fluid milk spoilage (e.g., spoilage due to post-pasteurization contamination with Gram-negative bacteria versus spoilage due to sporeforming bacteria). Microbiological data collected from shelf-life testing of 770 fluid milk samples were reviewed by 3 experts, who assigned milk samples into 3 distinct spoilage types, including (i) spoilage due to post-pasteurization contamination by Gram-negative bacteria, (ii) spoilage due to sporeformer contamination, and (iii) no microbial spoilage. This data set was then split into training, validation, and testing sets to develop a baseline model that can classify fluid milk spoilage. Additionally, 9 models based on subsets of data representing scenarios such as reducing the number of days or type of media needed for shelf-life testing were trained, validated, and subsequently tested based on the spoilage type classification accuracy. The baseline model, which included microbiological data from d 7, 14, and 21 of shelf-life, was validated with 91.7% classification accuracy and tested with 96.4% classification accuracy. Among the 9 models that represent different scenarios, the 2 models that included both standard plate counts and total Gram-negative counts (i) on d 14 and 21 of shelf-life and (ii) only on d 21 of shelf-life were both able to classify the spoilage types with 94.2% testing accuracy, suggesting a potential for optimizing shelf-life testing. Overall, our models have the potential to help milk processors (i) optimize their microbiological testing schemes to reduce costs and resources, (ii) identify the predominant spoilage pattern in a facility, and (iii) investigate the root-cause of spoilage to develop targeted interventions.