Abstract #M93

# M93
Detection of health problems by changes in milk estimated blood nonesterified fatty acids (NEFA) and milk fat, protein, and fatty acids.
Alex Pape*1, Heather M. Dann1, David M. Barbano2, Richard J. Grant1, 1William H. Miner Agricultural Research Institute, Chazy, NY, 2Department of Food Science, Northeast Dairy Food Research Center, Cornell University, Ithaca, NY.

The objective was to examine the relationship between changes in milk composition and the onset of ketosis or displaced abomasum (DA). The approach taken was to test the extent to which machine learning models could differentiate between milk samples from cows that went on to experience either ketosis or a DA and milk samples from cows that did not. Milk samples were analyzed with mid-infrared methodology. Milk-estimated blood NEFA (881.6 ± 304.6 μEq/L), de novo fatty acid (FA; 18.7 ± 3.7 g/100 g FA), preformed FA (50.6 ± 7.0 g/100 g FA), fat (5.1 ± 0.9%), the ratio of fat to protein (FTP; 1.4 ± 0.3), and the ratio of preformed FA to de novo FA (1.0 ± 2.9) were the specific milk composition features examined. Each milk sample from a cow with ketosis or DA was matched with ~10 milk samples from healthy cows. Each sample was from the same DIM and within 30 calendar days as the corresponding sample from a sick cow. Samples were drawn from a pool of 46,860 samples from 764 cows that were ≤21 DIM. A total of 1,436 samples were selected for the ketosis data set and 1,240 for the DA data set. Logistic regression and random forests models were used and evaluated with area under the ROC curve (AUC) from 10 replicates of 10-fold cross-validation. For ketosis and DA, milk-predicted blood NEFA, de novo FA, and preformed FA yielded AUC of ~0.8 from 5 to 0 d in advance of the event, with predictive power generally increasing up to an AUC of ~0.89 on the day of the event for logistic regression. This was found to be true for these 3 predictors both when considered individually and when combined. Fat, the FTP ratio, and the ratio of preformed FA to de novo FA yielded similar results with logistic regression, especially within 2 d of the event. Random forest results were more erratic but the overall predictive power found was not substantially different, indicating the selected variables are linearly related to the onset of health events. Overall, these results indicate that certain health events can be predicted with at least moderate accuracy several days in advance using changes in milk composition.

Key Words: transition cow, NEFA, milk fatty acid