Limited discriminatory power for predicting persistence of S. aureus intramammary infections using MALDI-TOF mass...

Staphylococcus aureus is a leading cause of intramammary infections (IMI) in Canadian dairy herds. These infections are frequently associated with elevated somatic cell counts (SCC), contributing to substantial economic losses through decreased milk quality and yield. Mass spectra generated by Matrix-Assisted Laser Desorption/Ionization Time-of-Flight Mass Spectrometry (MALDI-TOF MS) has more recently been explored to characterize genotypic traits of S. aureus. The objective of this longitudinal study was to evaluate the ability of machine learning algorithms to classify S. aureus IMI as persistent or short-term based exclusively on MALDI-TOF spectral features. As part of the National Cohort of Dairy Farms, milk samples were collected weekly from 15 lactating cows in 91 Canadian dairy herds over the period between 2007 and 2008. From over 137,000 quarter-level samples, we selected 1,602 quarters with at least one isolation of S. aureus. Persistent IMI were defined as ≥ 5 S. aureus-positive samples per quarter; short-term IMI were defined as a single S. aureus-positive sample preceded and followed by 2 negative samples. A total of 370 isolates were included in the study (222 persistent, 148 short-term). Spectra generated by MALDI-TOF MS were pre-processed, yielding 132 reference peaks from the training data set, which were then used to construct feature matrices for both the training and validation sets. Unsupervised analyses (PCA, k-means) revealed partial clustering but substantial overlap between IMI outcomes. Supervised models (Random Forest and Binary Discriminant Analysis) were used to classify isolates. The final random forest model achieved 62.1% accuracy (95% CI: 55.0-68.7%) on the validation set. In addition, a Kappa of 0.06 indicated poor agreement between predicted and true classes. For predicting a persistent infection, sensitivity was high at 99.2% (95% CI: 95.5-100.0%), but specificity was very low (0.06%, 95% CI: 0.02-0.1%), with an ROC of 0.53, suggesting no discriminative power. These findings suggested that MALDI-TOF mass spectra alone were insufficient to accurately predict persistent S. aureus IMI. In addition, the historical nature of the isolates used in this study should be considered when interpreting the applicability of these findings to contemporary S. aureus populations.