Microbiota-based prediction of Campylobacter contamination in broiler carcasses using random forest

Abstract Background Campylobacter is the leading cause of zoonotic disease in Europe, with poultry meat being the main source of human infection. Despite control measures, high levels of Campylobacter on broiler meat remain a major public health concern. Recent studies suggest that carcass microbiota may influence pathogen contamination, although this relationship remains largely unexplored for Campylobacter . Thus, the aim of this study was to determine whether the Campylobacter contamination status on broiler carcasses can be predicted from microbiota profiles at the end of slaughter and after refrigerated storage, using a tree-based machine learning approach (Random Forest). Bacterial taxa identified as predictors could serve as indirect microbiome-based indicators of Campylobacter contamination on broiler carcasses. Methods Random Forest models were developed using microbiological data from 314 broiler carcass samples. These models integrated culture-based quantification of Campylobacter and bacterial community characterised by 16S rRNA gene metabarcoding. Two models were built using data collected after slaughter (D0) and after refrigerated storage (D7). Samples were classified as “Positive” when the Campylobacter concentration exceeded 1 log₁₀ CFU/mL, and “Negative” when it was below this threshold. Results Random Forest models achieved robust performance: accuracy was 92.8% at D0 and 99.6% at D7, and AUC was 1 at both D0 and D7. The most predictive taxa at D0 included ASVs related to Faecalibacterium unknown species and Bacteroides spp. , all of which showed a marginal positive effect in the Partial Dependence Plots, indicating an increased probability of Campylobacter -positive status. In contrast, other important taxa at D0, such as Lactobacillus unknown species, Psychrobacter sp. and Limosilactobacillus unknown species, exhibited a negative marginal effect on the model-predicted outcome. At D7, different bacterial species were identified as key predictors, including Pseudomonas spp., Flavobacterium spp., Psychrobacter sp., Lactococcus raffinolactis , Aeromonas sp. and Acinetobacter spp. Of these, Lactococcus raffinolactis , Aeromonas sp., and Pseudomonas spp. exhibited a positive marginal effect on the model’s predicted outcome, while Flavobacterium spp. and Acinetobacter spp. showed a negative marginal effect. Conclusion Our findings highlight the potential of machine learning models to identify microbiome-based indicators, paving the way for targeted monitoring strategies and improved food safety.