Abstract :
[en] Among the dairy sector's current concerns, the early detection of animal health disorders is a
complex challenge as it includes different diseases. This multidimensionality explains why
disease detection is often studied separately and, due to financial and ethical issues, using
small-scale datasets. Several studies were conducted in the past using the milk mid-infrared
(MIR) spectra, for instance, to detect mastitis, lameness or to quantify the contents of citrate,
β-hydroxybutyrate (BHB) or acetone in milk. To solve this issue and the small scale data size,
we considered a holistic approach using traits obtained from milk recording to detect animal
health disorders: milk yield, the somatic cell count and 27 MIR predictions related to the milk
composition and animal health status. From 740,054 records collected from first parity
Holstein cows in the Southern part of Belgium, we performed repeated unsupervised
learnings. The obtained clustering divided the records into five groups. Significant differences
of feature means were found between groups, suggesting that one group was related to
mastitis and a second group to metabolic disorders. A validation from 87 milk and blood
reference records obtained through the Interreg European project GplusE confirmed this
interpretation. Moreover, after using a principal components analysis performed on the used
features, it appeared that the first and fourth principal components (PC) were strongly related
to the two discovered groups of sick animals. From reference values, the first PC had
correlations of -0.68 with blood BHB, -0.70 with blood Non-Esterified Fatty Acids, 0.61 with
blood Glucose and -0.46 with milk isocitrate. On the other hand, the fourth PC had correlations
of 0.51 with milk N-acetyl-β-D-glucosaminidase and 0.55 with milk lactate dehydrogenase.
Those results suggest that the obtained PCs reflect directly main health disorders and could
be used to monitor dairy farms on large scale data.
