96 Association between metabolic diseases and fertility of high-yielding dairy cows in a transition management facility using survival analysis and machine-learning models

Abstract

This study aimed to evaluate the association between individual and multiple metabolic diseases (MD and MD+) diagnosed during the transition period (± 3 wk relative to calving) and the probability of pregnancy until 210 days in milk (DIM) in Holstein-Friesian dairy cows. Disease and reproductive data from a dairy herd with 1946 calvings (n = 542 primiparous and n = 1404 multiparous cows) were analysed using a 1-year cohort. The recorded MD were milk fever, ketosis, displaced abomasum, retained placenta, metritis, twinning, and clinical mastitis. The association between the 210-DIM pregnancy risk and the MD was evaluated as MD cows (uncomplicated cases) v. MD+ cows (complicated cases) v. healthy cows (3 groups of cows). Univariable survival models were used to analyse the association of MD and MD+ with pregnancy until 210 DIM, accounting for parity. Univariable Cox proportional hazard models were used to quantify the relative risk of pregnancy per day. A hierarchically ordered decision tree and a random forest model were built to explore the importance of MD and parity on the pregnancy risk within the first 210 DIM. Parity affected the 210-DIM pregnancy risk (P < 0.001); therefore, all further analyses were stratified by parity. The incidence of MD and MD+ for primiparous and multiparous cows were 29 (n = 159) and 9% (n = 48), and 23 (n = 317) and 11% (n = 160), respectively. The overall 210-DIM pregnancy risk was 77% (n = 415) for primiparous cows and 62% (n = 879) for multiparous cows. Among healthy cows (no MD) the 210-DIM pregnancy risk was 80% (n = 269) for primiparous cows and 82% (n = 537) for multiparous cows. Conversely, the 210-DIM pregnancy risk for cows with MD or MD+ were 73 (n = 116) and 63% (n = 30) for primiparous and 48 (n = 152) and 46% (n = 74) for multiparous cows, respectively. Using the healthy cows as the reference, the 210-DIM hazard ratios for conception were 0.8 for MD [95% confidence interval (CI) = 0.6-1.0; P = 0.05] and 0.5 for MD+ (95% CI = 0.4-0.8; P = 0.005) for primiparous cows and 0.5 for MD (95% CI = 0.4-0.6; P < 0.001) and 0.4 for MD+ (95% CI = 0.3-0.6; P < 0.001) for multiparous cows. Parity had profound effect on the 210-DIM pregnancy risk. The hazard ratio for conception was reduced when a MD was complicated with another MD (MD+) in both primiparous and multiparous cows. Both the decision tree and random forest analysis also indicated that parity was the most influential variable reducing fertility among cows, followed by (in order of magnitude of effect) milk fever, displaced abomasum, ketosis, and clinical mastitis. Including multiple disease interactions into multivariable Cox proportional hazard models are highly likely to violate the proportional hazards assumption due to complex disease interactions. Machine-learning models represent a valid alternative to accommodate large datasets in the presence of missing values and intricate dependencies among explanatory variables.