Predicting dyscalcemia status in early-lactation multiparous Holstein cows using milk weight and constituent analysis from a single milking at 4 days in milk.

Many multiparous cows struggle to adapt to the challenges of the early postpartum period. Dyscalcemia, a condition defined by low blood calcium concentrations at 4 DIM and associated with suboptimal performance across a spectrum of epidemiologically important outcomes (health, productivity, and reproductive success), can be a useful indicator that maladaptive phenotypes are developing in early postpartum dairy cows. Identifying dyscalcemic cows, though theoretically useful from a management perspective, is not logistically viable for commercial dairy farms due to the costs and labor that would be involved in the collection and analysis of samples. Furthermore, timely methods of analysis are lacking. Therefore, our objective in this cross-sectional study was to develop a predictive model for establishing dyscalcemia status by applying machine learning approaches to milk weights and milk constituent data predicted using Fourier-transform mid-infrared spectroscopy (FTIR) from a single milking at 4 DIM. We hypothesized that such a model would have adequate diagnostic characteristics. To test this hypothesis, we collected blood, milk weights, and proportional milk samples from 542 multiparous Holsteins on 5 commercial dairy farms in central New York at 4 DIM. Blood was analyzed for serum total calcium concentration and milk was subjected to FTIR analysis from which constituent data were predicted. Cows were diagnosed as having dyscalcemia if they had serum total calcium concentration ≤2.2 mmol/L at 4 DIM, and as eucalcemic if their serum total calcium concentrations were >2.2 mmol/L at this time. Using milk yield data and the concentrations of anhydrous lactose, true protein, fat, and fatty acid groups, including de novo, mixed, and preformed, all measured in g/100 g milk, as well as milk urea nitrogen (mg/100 g milk), and milk ketone bodies (BHB and acetone; mmol/L) we fit and cross validated random forest models stratified by parity group (2, 3, and ≥4) and farm, for the prediction of dyscalcemia status, our main outcome of interest. We found that on average our models performed favorably with an area under the receiver operating characteristic curve of 0.95 (95% CI: 0.86-1.00), accuracy of 0.90 (95% CI: 0.81-0.98), sensitivity of 0.85 (95% CI: 0.64-1.00), specificity of 0.91 (95% CI: 0.84-1.00), positive predictive value of 0.71 (95% CI: 0.32-1.00) and negative predictive value of 0.96 (95% CI: 0.89-1.00). The data providing the most valuable information to our models were milk weight, and concentrations of lactose and protein. These findings, though limited to a single geographic region, time of day, milking schedule, and season, support the concept that machine learning approaches combined with milk constituent data could become a valuable tool for discriminating between dyscalcemic cows and their eucalcemic counterparts in the early postpartum period.