Advertisement

Bayesian estimation of sensitivity and specificity of a milk pregnancy-associated glycoprotein ELISA test for pregnancy diagnosis between 23 and 27 days after insemination in Holstein dairy cows

Open AccessPublished:July 01, 2022DOI:https://doi.org/10.3168/jds.2022-21905

      ABSTRACT

      Pregnancy diagnosis using pregnancy-associated glycoprotein (PAG) ELISA technology in blood or milk samples is validated from 28 d after insemination in dairy cows. The objective of this study was to estimate the sensitivity (Se) and specificity (Sp) of a commercial milk PAG-based ELISA in Holstein dairy cows between 23 and 27 d after insemination. Milk samples (n = 268) from 257 Holstein dairy cows 23 to 27 d after AI were submitted for PAG ELISA testing. Pregnancy status was confirmed by either a second milk PAG ELISA test conducted between 28 and 50 d after insemination (n = 200) or transrectal ultrasonography performed between 28 and 59 d after insemination (n = 68). A Bayesian latent class model was used to compare the paired results from the test at 23 to 27 d after AI test to the reference test. The latent class model typically used for comparing 2 or more imperfect tests was extended to include the possibility of pregnancy loss between the 23 to 27 d test and the reference test. Informative priors for the probability of pregnancy loss, and for the Se and Sp of the PAG and ultrasonography reference tests were obtained from the scientific literature. Estimated median Se and Sp of the PAG ELISA test conducted between 23 and 27 d after AI were 0.98 (95% credible interval 0.93 to 1.0) and 0.98 (0.89 to 1.0), respectively, when using a standardized corrected optical density threshold of 0.15. Although the accuracy of the test under investigation was excellent, more data will be needed to confirm the optimal diagnostic cut point for PAG in milk for early pregnancy diagnosis in this time window. The optimal timing of pregnancy diagnosis will depend on herd-specific logistics and the action to be taken to re-inseminate nonpregnant cows.

      Key words

      INTRODUCTION

      Early pregnancy diagnosis is a critical component of a reproductive management program because rapid accurate identification of nonpregnant cows will contribute to minimizing time to re-insemination. Transrectal ultrasonography (US) is widely used for early pregnancy diagnosis in dairy herds around the world, but there is a growing use of milk pregnancy-associated glycoproteins (PAG) ELISA tests. Although some producers employ blood PAG ELISA testing, milk samples are considered more convenient by many farmers. The DHI samples collected at milk recording or hand-stripped samples collected at the producer's convenience are the testing options presently available. Pregnancy-associated glycoproteins are produced by binucleate trophoblast cells during pregnancy before reaching the maternal circulation (
      • Green J.A.
      • Xie S.
      • Roberts R.M.
      Pepsin-related molecules secreted by trophoblast.
      ). Levels of PAG above an established cut point in a blood or a milk sample are indicative of pregnancy, and the diagnostic accuracy of this technology has been evaluated in many studies (
      • Lawson B.C.
      • Shahzad A.H.
      • Dolecheck K.A.
      • Martel E.L.
      • Velek K.A.
      • Ray D.L.
      • Lawrence J.C.
      • Silvia W.J.
      A pregnancy detection assay using milk samples: Evaluation and considerations.
      ;
      • Karen A.
      • Sousa N.M.D.
      • Beckers J.-F.
      • Bajcsy Á.C.
      • Tibold J.
      • Mádl I.
      • Szenci O.
      Comparison of a commercial bovine pregnancy-associated glycoprotein ELISA test and a pregnancy-associated glycoprotein radioimmunoassay test for early pregnancy diagnosis in dairy cattle.
      ;
      • Ricci A.
      • Carvalho P.D.
      • Amundson M.C.
      • Fourdraine R.H.
      • Vincenti L.
      • Fricke P.M.
      Factors associated with pregnancy-associated glycoprotein (PAG) levels in plasma and milk of Holstein cows during early pregnancy and their effect on the accuracy of pregnancy diagnosis.
      ;
      • Commun L.
      • Velek K.
      • Barbry J.-B.
      • Pun S.
      • Rice A.
      • Mestek A.
      • Egli C.
      • Leterme S.
      Detection of pregnancy-associated glycoproteins in milk and blood as a test for early pregnancy in dairy cows.
      ;
      • Mercadante P.M.
      • Ribeiro E.S.
      • Risco C.
      • Ealy A.D.
      Associations between pregnancy-associated glycoproteins and pregnancy outcomes, milk yield, parity, and clinical diseases in high-producing dairy cows.
      ).
      One major challenge when evaluating the accuracy of a PAG ELISA test for early pregnancy diagnosis is the absence of a gold-standard test. From approximately 28 d, visualization of an embryo with a heartbeat by US may be considered the reference method but it is imperfect (
      • Fricke P.M.
      • Ricci A.
      • Giordano J.O.
      • Carvalho P.D.
      Methods for and implementation of pregnancy diagnosis in dairy cows.
      ). Furthermore, true status may change quickly because losses are common in early pregnancy. An alternative approach in this situation is the use of Bayesian latent class model (LCM) methodology. A study (
      • Dufour S.
      • Durocher J.
      • Dubuc J.
      • Dendukuri N.
      • Hassan S.
      • Buczinski S.
      Bayesian estimation of sensitivity and specificity of a milk pregnancy-associated glycoprotein-based ELISA and of transrectal ultrasonographic exam for diagnosis of pregnancy at 28–45 days following breeding in dairy cows.
      ) using Bayesian LCM analysis showed that US and a commercial milk PAG-based test were highly accurate from 28 to 45 d following insemination in lactating dairy cows. The median and 95% credible intervals (CrI) for sensitivity (Se) and specificity (Sp) were 0.96 (0.91–1.00) and 0.99 (0.97–1.00), respectively, for US and were 0.99 (0.98–1.00) and 0.95 (0.89–1.0), respectively, for the milk PAG test when using the manufacturer's recommended thresholds [i.e., cows with milk standardized corrected optical density (OD) <0.10 identified as open, those with standardized corrected OD ≥0.25 identified as pregnant, and those with OD ≥0.10 and <0.25 having an undetermined status].
      • Fosgate G.T.
      • Motimele B.
      • Ganswindt A.
      • Irons P.C.
      A Bayesian latent class model to estimate the accuracy of pregnancy diagnosis by transrectal ultrasonography and laboratory detection of pregnancy-associated glycoproteins in dairy cows.
      showed similar results for the same commercial milk PAG test using Bayesian LCM methodology with estimated Se and Sp (95% CrI) of 0.99 (0.98–1.00) and 0.93 (0.90–0.96), respectively, in dairy cows 28 to 35 d after insemination in South Africa.
      Commercial PAG testing products available in Canada for either blood or milk samples are labeled for use from 28 or 29 d after AI in dairy cows. Nevertheless, some pregnant cows have serum PAG levels significantly higher than open cows as early as 22 to 24 d following insemination (
      • Friedrich M.
      • Holtz W.
      Establishment of an ELISA for measuring bovine pregnancy-associated glycoprotein in serum or milk and its application for early pregnancy detection.
      ).
      • Green J.A.
      • Parks T.E.
      • Avalle M.P.
      • Telugu B.P.
      • McLain A.L.
      • Peterson A.J.
      • McMillan W.
      • Mathialagan N.
      • Hook R.R.
      • Xie S.
      • Roberts R.M.
      The establishment of an ELISA for the detection of pregnancy-associated glycoproteins (PAGs) in the serum of pregnant cows and heifers.
      followed the serum PAG levels of 42 pregnant cows and showed that PAG became detectable as early as 22 d after insemination in a small percentage of the pregnant cows, in most of them by d 25, and in all pregnant cows by d 28. Because earlier pregnancy diagnosis may allow for earlier rebreeding of open cows, the earliest accurate diagnosis is desirable to allow for a range of options to achieve timely and fertile re-insemination. A series of experiments was conducted to measure the accuracy of an in-house PAG ELISA test on blood samples at d 24 of gestation.
      • Reese S.T.
      • Pereira M.H.C.
      • Edwards J.L.
      • Vasconcelos J.L.M.
      • Pohler K.G.
      Pregnancy diagnosis in cattle using pregnancy associated glycoprotein concentration in circulation at day 24 of gestation.
      used Gir × Holstein cows and concluded that their assay was not accurate enough at this stage of the pregnancy.
      • Oliveira Filho R.V.
      • Franco G.A.
      • Reese S.T.
      • Dantas F.G.
      • Fontes P.L.P.
      • Cooke R.F.
      • Rhinehart J.D.
      • Thompson K.W.
      • Pohler K.G.
      Using pregnancy associated glycoproteins (PAG) for pregnancy detection at day 24 of gestation in beef cattle.
      used Bos taurus beef cattle in the United States at d 24 of pregnancy. Using different cutoff values for heifers and cows, Se and Sp were 0.86 and 0.90, respectively, for heifers and 0.73 and 0.88 for cows. The Se of a commercial milk PAG test performed at 25 and 32 d after AI was evaluated based on US of 48 lactating Holstein cows bearing singleton pregnancies at 32 d of gestation (
      • Ricci A.
      • Carvalho P.D.
      • Amundson M.C.
      • Fourdraine R.H.
      • Vincenti L.
      • Fricke P.M.
      Factors associated with pregnancy-associated glycoprotein (PAG) levels in plasma and milk of Holstein cows during early pregnancy and their effect on the accuracy of pregnancy diagnosis.
      ). At 25 d after insemination, 25 of the 48 cows were classified as pregnant, 5 as nonpregnant and 18 inconclusive by the PAG test. At 32 d after insemination, 47 out of 48 pregnant cows were classified as pregnant, and the remaining one had an inconclusive result based on PAG in milk. The Se of the commercial milk PAG test based on the pregnant, and nonpregnant classifications only was 0.83 and 1.00 at 25 and 32 d postinsemination, respectively. In addition to the low Se at 25 d, another concern was the high proportion of inconclusive results at 25 compared with 32 d after insemination.
      Estimating Sp of a pregnancy test performed before 28 d after AI is a challenge. Because failure to visualize an embryo on US is not definitive for nonpregnancy until at least 28 d, another test cannot be applied at the same time, so the only comparator is another test applied later. Pregnancy loss is approximately 12% between 28 and 60 d (
      • Wiltbank M.C.
      • Baez G.M.
      • Garcia-Guerra A.
      • Toledo M.Z.
      • Monteiro P.L.J.
      • Melo L.F.
      • Ochoa J.C.
      • Santos J.E.P.
      • Sartori R.
      Pivotal periods for pregnancy loss during the first trimester of gestation in lactating dairy cows.
      ), resulting in a possible underestimation of the Sp of the test in early gestation.
      Our objective was to estimate the Se and Sp of a commercial milk PAG-based ELISA [Alertys Milk Pregnancy Test (formerly named Idexx Milk Pregnancy Test), Idexx Laboratories] in Holstein dairy cows 23 to 27 d following insemination, accounting for the imperfection of the reference tests and the possibility of pregnancy loss between the 23 to 27 d test and later reference test. We used Bayesian LCM to achieve this objective using imperfect tests and the possibility of a change in status between tests.

      MATERIALS AND METHODS

      Study Design

      A longitudinal prospective study was conducted to determine the accuracy of a commercial milk PAG ELISA test applied between 23 and 27 d after insemination. Guidelines from the Standards for the Reporting of Diagnostic Accuracy that Use Bayesian LCM statement (
      • Kostoulas P.
      • Nielsen S.S.
      • Branscum A.J.
      • Johnson W.O.
      • Dendukuri N.
      • Dhand N.K.
      • Toft N.
      • Gardner I.A.
      STARD-BLCM: Standards for the Reporting of Diagnostic accuracy studies that use Bayesian Latent Class Models.
      ) were used for reporting. For that study, cows from commercial dairy farms were used. Regarding animal care, this was an observational study, with no additional intervention beyond what was currently done on these farms, with the exception of the collection of 1 milk sample. During the study, all farmers were required to follow the guidelines of the 2009 Code of Practice for the Care and Handling of Dairy Cattle (National Farm Animal Care Council, https://www.nfacc.ca/pdfs/codes/dairy_code_of_practice.pdf). To ensure reproducibility of our analyses, the assembled data set, all the R scripts used for analyses, and additional descriptive analyses are available as Supplemental Files S1 to S4 (https://doi.org/10.5683/SP3/EEQMRL;
      • Durocher J.
      • Moore R.K.
      • Albaaj A.
      • Leblanc S.
      • Dufour S.
      Bayesian estimation of sensitivity and specificity of a commercial milk pregnancy-associated glycoprotein ELISA test between 23 and 27 days after breeding in dairy cows.
      ).

      Animals and Herds

      The target population was lactating Holstein cows from commercial dairy herds in Québec, Canada. A convenience sample of one tiestall and one freestall herd was selected at the beginning of the study. These 2 herds were already using the commercial milk PAG ELISA test under investigation on a regular basis for pregnancy diagnosis starting 28 d following insemination and were enrolled on DHI. These producers were used to the sampling and shipping procedures associated with PAG ELISA testing from hand-stripped samples. Veterinary herd health visits were scheduled monthly and all information regarding US pregnancy diagnosis and other reproductive events were available. Two other tiestall herds also submitted a limited number of samples later in the project. The study period for these 4 herds ran from November 2017 to October 2018. Data from a few cows (n = 8) from a previous study (
      • Dufour S.
      • Durocher J.
      • Dubuc J.
      • Dendukuri N.
      • Hassan S.
      • Buczinski S.
      Bayesian estimation of sensitivity and specificity of a milk pregnancy-associated glycoprotein-based ELISA and of transrectal ultrasonographic exam for diagnosis of pregnancy at 28–45 days following breeding in dairy cows.
      ) that met the criteria of the current study protocol (i.e., having both a first PAG ELISA test conducted between 23 and 27 d postinsemination and a subsequent PAG ELISA test or US exam 28 to 59 d after insemination) were also added to the data set. These samples were collected from 3 tiestall and one freestall herds from November 2014 to January 2015. No other selection criteria were used to select breeding events included in the study.

      Testing Methods

      Cows between 23 and 27 d following insemination were enrolled in the study. A 15-mL quarter-milk sample (hand-stripped sample from one quarter) was collected by the producer. A bronopol tablet was added to the milk sample for preservation, as done with DHI milk samples. The samples were sent to an ISO 17025 certified laboratory (Lactanet, Sainte-Anne-de-Bellevue, QC, Canada) following the shipping procedures already in place for the Lactanet commercial PAG ELISA testing service. A second milk sample was collected for pregnancy confirmation, ≥28 d following insemination, but no later than 59 d, and following the same sampling and shipping procedures. Milk samples were collected and shipped once a week. The exact day of sampling was at the convenience of the farmer and their schedule for shipping milk samples for pregnancy tests. The producers also had the possibility to replace the second milk sampling by a pregnancy diagnosis by US by the herd veterinarian if it was more convenient for them. The US exam also had to be conducted between 28 and 59 d after insemination.
      Samples from cows tested between 23 and 27 d following insemination followed the same process as other samples collected on the same day. Results from samples between 23 and 27 d were reported to the producer as “inconclusive test” because they were sampled before 28 d after insemination. Producers were asked to wait until the test conducted between 28 and 59 d after insemination to act based on the pregnancy diagnosis.
      Milk samples were analyzed by 1 of 2 technicians involved in the PAG ELISA testing laboratory, who were blinded to the results of the other pregnancy tests. The technicians had previously received training on milk ELISA testing. Milk samples were analyzed within 24 h of receipt using a commercial ELISA kit for detection of PAG in milk (Alertys Milk Pregnancy test, Idexx laboratories). Manufacturer's guidelines were followed for the analyses. Briefly, 96-well microtiter plates were coated with an anti-PAG antibody. After 2 h of incubation, test samples were washed and captured PAG was detected with a PAG-specific antibody and horseradish peroxidase conjugate. Unbound conjugate was washed away and tetramethylbenzidine substrate was added to the wells. Color development is proportional to the amount of PAG present in the samples. Optical density, which varies depending on PAG concentration, was measured using the ELx808 Biotek absorbance microplate reader equipped with a single 450 nm filter and reference wavelength of 620 nm to 650 nm (REF). From corrected OD values (450–REF), the mean of the negative controls was subtracted to compute standardized corrected OD values for each sample. These final values were used for interpretation. For the PAG ELISA test at 23 to 27 d we evaluated different dichotomous thresholds for classification of pregnancy, from a standardized corrected OD of 0.05 to 0.40 in increments of 0.05. Accordingly, for this analysis, no results were classified as inconclusive.
      For cows that had their pregnancy status confirmed by a milk PAG ELISA test 28 to 59 d after AI, those with milk standardized corrected OD <0.25 were classified as nonpregnant, and those with standardized corrected OD ≥0.25 were classified as pregnant based on
      • Dufour S.
      • Durocher J.
      • Dubuc J.
      • Dendukuri N.
      • Hassan S.
      • Buczinski S.
      Bayesian estimation of sensitivity and specificity of a milk pregnancy-associated glycoprotein-based ELISA and of transrectal ultrasonographic exam for diagnosis of pregnancy at 28–45 days following breeding in dairy cows.
      .
      Interpretation of the US exam was done by the veterinarian using visualization of amniotic fluid, the fetus, or a fetal heartbeat. Only the final interpretation (i.e., pregnant or open) of the US exam was available. Whether a fetal heartbeat was observed or not was not recorded. Type, brand, and model of US units varied among practitioners. Cows with an undetermined US status were later re-evaluated by the veterinarian and only the result and date of this second evaluation was used in our subsequent analyses.

      Statistical Analyses

      For each 23 to 27 d PAG ELISA threshold, we generated the cross-tabulated results between the 23 to 27 d PAG test and the reference test. Cross-tabulated results were generated separately for cows having a PAG ELISA confirmation ≥28 d and cows confirmed using US. Thus, for each potential threshold, two 2-by-2 tables were made.
      A Bayesian LCM for 2 populations in which the same diagnostic test under investigation (PAG ELISA at 23 to 27 d) is used in both populations, but different reference tests are used (PAG ELISA ≥28 d or US ≥28 d) was developed. In this LCM, the theoretical latent class that we were trying to model was the presence of a viable pregnancy, when the first diagnostic test (PAG ELISA 23–27 d after insemination) was applied. We expanded this LCM to account for the probability of pregnancy loss between the 23 to 27 d PAG ELISA and reference tests as illustrated in Figure 1. Briefly, we added to the matrices of probabilities typical of these LCM, the probability of pregnancy loss (PL) or of pregnancy retention (1 − PL) between tests to the cows that were truly pregnant at 23 to 27d after insemination. Using this layout, one can generate a set of equations explaining all potential combinations of tests results (for a detailed description, see Supplemental File S1). In these equations, setting PL to zero would yield the typical LCM for diagnostic test validation (i.e., explicitly assuming that there is no change of health status between tests).
      Figure thumbnail gr1
      Figure 1Distribution of potential outcomes of 2 tests, a and b (Ta+ or Ta− and Tb+ or Tb−) and their respective probabilities, as a function of the true status (in this case, pregnant or not pregnant; True+ or True−) and of the tests' sensitivities (Sea and Seb) and specificities (Spa and Spb) for 2 conditionally independent tests, when the individual's status can change from positive to negative during the interval between the 2 tests. True prevalence of cows having the status of interest (here, pregnancy) when the first test (Ta) is conducted is denoted as Pi, and probability of a change in status (probability of pregnancy loss) between the initial test and subsequent reference test (Tb) is denoted as PL.
      As proposed by
      • Dendukuri N.
      • Joseph L.
      Bayesian approaches to modeling the conditional dependence between multiple diagnostic tests.
      , we further expanded the LCM to relax the assumption of conditional independence between tests for cows for which a PAG ELISA was used as reference test. For simplicity, we did not try to relax the assumption of conditional independence between the PAG ELISA test at 23 to 27 d and the US test ≥28 d, because these tests rely on different biological processes (measurement of a marker in milk vs. visualization of anatomical structures).
      Finally, a last step was added to compute the negative (NPV) and positive predictive values (PPV) using the prevalence of pregnancy at 23 to 27 d after insemination estimated from the model in the population of cows in which PAG ELISA ≥28 d was used as reference test. The coding and data set for the LCM developed are provided in Supplemental Files S1 and S2, respectively.

      Prior Distributions for the Unknown Parameters

      For Se and Sp of the reference tests we used the results from
      • Dufour S.
      • Durocher J.
      • Dubuc J.
      • Dendukuri N.
      • Hassan S.
      • Buczinski S.
      Bayesian estimation of sensitivity and specificity of a milk pregnancy-associated glycoprotein-based ELISA and of transrectal ultrasonographic exam for diagnosis of pregnancy at 28–45 days following breeding in dairy cows.
      to generate informative prior distributions. Briefly, for Se and Sp of the PAG ELISA ≥28 d, we used prior distributions corresponding to a mode (fifth percentile) of 0.99 (0.98) and 0.96 (0.91), respectively [corresponding to a β(100, 2) and a β(92, 5) distribution, respectively]. For Se and Sp of the US exam ≥28 d, we used prior distributions corresponding to a mode (fifth percentile) of 0.94 (0.90) and 0.95 (0.90), respectively [corresponding to a β(100, 7) and a β(100, 6) distribution, respectively]. For the Se and Sp of the PAG ELISA at 23 to 27 d (the test under investigation) uniform β(1, 1) prior distributions were used, thus explicitly indicating that all values between 0 and 1 were equally probable.
      To generate prior distributions for prevalence of pregnancy between 23 and 27 d after insemination, we assembled the data from all PAG tests 28 d after insemination (n = 293,104) analyzed between 2016 and 2021 by Lactanet, Canada (the Canadian DHI company). Using these data, we observed annual proportions of tests classified pregnant at 28 d postinsemination between 0.55 and 0.59. Therefore, we chose a distribution centered on 0.55, but relatively diffuse (i.e., with fifth percentile at 0.35), because the period (28 d after insemination) did not correspond exactly to the targeted period (23–27 d). This distribution was described as a β(9, 8) distribution. For the probability of pregnancy loss between the initial test and the reference test, we searched the scientific literature for studies reporting the probability of pregnancy loss in dairy cattle during the stage of gestation when our tests were conducted. We found 5 relevant sources of information on probability of pregnancy loss. A review article (
      • Santos J.E.P.
      • Thatcher W.W.
      • Chebel R.C.
      • Cerri R.L.A.
      • Galvão K.N.
      The effect of embryonic death rates in cattle on the efficacy of estrus synchronization programs.
      ) presented results from different studies (
      • Vasconcelos J.L.M.
      • Silcox R.W.
      • Lacerda J.A.
      • Pursley J.R.
      • Wiltbank M.C.
      Pregnancy rate, pregnancy loss, and response to head stress after AI at 2 different times from ovulation in dairy cows.
      ;
      • Moreira F.
      • Orlandi C.
      • Risco C.A.
      • Mattos R.
      • Lopes F.
      • Thatcher W.W.
      Effects of presynchronization and bovine somatotropin on pregnancy rates to a timed artificial insemination protocol in lactating dairy cows.
      ;
      • Santos J.E.
      • Thatcher W.W.
      • Pool L.
      • Overton M.W.
      Effect of human chorionic gonadotropin on luteal function and reproductive performance of high-producing lactating Holstein dairy cows.
      ,
      • Santos J.E.P.
      • Bartolome J.A.
      • Cerri R.L.A.
      • Juchem S.O.
      • Hernandez O.
      • Trigg T.
      • Thatcher W.W.
      Effect of a deslorelin implant in a timed artificial insemination protocol on follicle development, luteal function and reproductive performance of lactating dairy cows.
      ;
      • Silke V.
      • Diskin M.G.
      • Kenny D.A.
      • Boland M.P.
      • Dillon P.
      • Mee J.F.
      • Sreenan J.M.
      Extent, pattern and factors associated with late embryonic loss in dairy cows.
      ;
      • Chebel R.C.
      • Santos J.E.P.
      • Cerri R.L.A.
      • Galvão K.N.
      • Juchem S.O.
      • Thatcher W.W.
      Effect of resynchronization with GnRH on day 21 after artificial insemination on pregnancy rate and pregnancy loss in lactating dairy cows.
      ;
      • Galvão K.N.
      • Santos J.E.P.
      • Juchem S.O.
      • Cerri R.L.A.
      • Coscioni A.C.
      • Villaseñor M.
      Effect of addition of a progesterone intravaginal insert to a timed insemination protocol using estradiol cypionate on ovulation rate, pregnancy rate, and late embryonic loss in lactating dairy cows.
      ;
      • Cerri R.L.A.
      • Juchem S.O.
      • Chebel R.C.
      • Rutigliano H.M.
      • Bruno R.G.S.
      • Galvão K.N.
      • Thatcher W.W.
      • Santos J.E.P.
      Effect of fat source differing in fatty acid profile on metabolic parameters, fertilization, and embryo quality in high-producing dairy cows.
      ). Unpublished data from Toledo were also available in
      • Wiltbank M.C.
      • Baez G.M.
      • Garcia-Guerra A.
      • Toledo M.Z.
      • Monteiro P.L.J.
      • Melo L.F.
      • Ochoa J.C.
      • Santos J.E.P.
      • Sartori R.
      Pivotal periods for pregnancy loss during the first trimester of gestation in lactating dairy cows.
      and finally data from 3 primary studies (
      • Romano J.E.
      • Thompson J.A.
      • Forrest D.W.
      • Westhusin M.E.
      • Tomaszweski M.A.
      • Kraemer D.C.
      Early pregnancy diagnosis by transrectal ultrasonography in dairy cattle.
      ;
      • Fricke P.M.
      • Ricci A.
      • Giordano J.O.
      • Carvalho P.D.
      Methods for and implementation of pregnancy diagnosis in dairy cows.
      ;
      • Middleton E.L.
      • Pursley J.R.
      Short communication: Blood samples before and after embryonic attachment accurately determine non-pregnant lactating dairy cows at 24 d post-artificial insemination using a commercially available assay for pregnancy-specific protein B.
      ) were also available. The period investigated varied from one study to another, but, in general, focused on 27 to 45 d of gestation. For this period, we conducted a random effect meta-analysis including 6,458 records from 17 studies. The heterogeneity of the data set was high (I2 = 88%; Q statistics: χ2 = 138.60, degrees of freedom = 16, P < 0.0001) and the meta-analysis showed a summary estimate of 0.11 (95% CrI = 0.09–0.14) for pregnancy loss. We found only one study describing pregnancy loss between 45 and 60 d of gestation (
      • Vasconcelos J.L.M.
      • Silcox R.W.
      • Lacerda J.A.
      • Pursley J.R.
      • Wiltbank M.C.
      Pregnancy rate, pregnancy loss, and response to head stress after AI at 2 different times from ovulation in dairy cows.
      ) and it was estimated at 0.06. Moreover, a minority of cows in our study had their reference test >45 d after insemination. For simplicity, we described the probability of pregnancy loss as a β distribution corresponding to a mode of 0.10 and 95th percentile at 0.20 [corresponding to a β(6, 43)].
      Prior distributions for the positive and negative covariance between the PAG ELISA at 23 to 27 and ≥28 d were described using the natural bounds of these covariance terms as described by
      • Dendukuri N.
      • Joseph L.
      Bayesian approaches to modeling the conditional dependence between multiple diagnostic tests.
      . The prior distributions for all the unknown parameters are illustrated in Supplemental Files S1 to S4.

      Implementation of the LCM

      The described LCM were run, using the described prior distributions, and for each of the potential thresholds (from 0.05 to 0.40 in increments of 0.05) of PAG ELISA 23 to 27 d after AI using the corresponding 2-by-2 tables with the test results. We used the package R2OpenBUGS (version 3.2–3.2.1;
      • Sturtz S.
      • Ligges U.
      • Gelman A.
      R2WinBUGS: A package for running WinBUGS from R.
      ) in R (version 4.0.4) using the RStudio platform (version 1.4.1106; https://r-project.org) for all analyses. This package executed the analyses through the OpenBUGS software (version 3.2.3).
      Each LCM was run with 3 Markov chains with different sets of initial values for the unknown parameters, and with a burn-in period of 1,000 iterations. Following this burn-in period, the Monte Carlo simulation was run to generate 20,000 values/Markov chain to store, for each LCM, 60,000 values from the posterior distributions. Trace and autocorrelation plots were visually inspected to ensure chains' convergence and chains' behavior. The effective sample size was computed to ensure that a sufficiently large number of independent values were generated to describe the posterior distributions with sufficient precision. The R scripts used are available in Supplemental File S1.

      Sensitivity Analysis

      To evaluate the effect of the chosen informative priors we conducted a Se analysis in which the same LCM where run, but using perturbed and more diffuse priors, as described by
      • Johnson W.O.
      • Jones G.
      • Gardner I.A.
      Gold standards are out and Bayes is in: Implementing the cure for imperfect reference tests in diagnostic accuracy studies.
      . Briefly, we modified the mode for the Se and Sp of the reference tests (PAG ELISA ≥28 d and US ≥ 28d) so that they would be 10 percentage points lower than initially specified and so that the fifth percentile would be 10 percentage points from the mode. The corresponding β distributions, along with the R script used for these analyses, are in Supplemental File S3.
      Finally, to investigate the effect of accounting or not for pregnancy loss between tests, we ran the initial LCM, with all the informative prior distributions, but with a probability of pregnancy loss of exactly 0 (thus explicitly stating that no change in pregnancy status would occur between tests). The corresponding R script is also in Supplemental File S4.

      Sample Size Estimation

      The method described by
      • Georgiadis M.P.
      • Johnson W.O.
      • Gardner I.A.
      Sample size determination for estimation of the accuracy of two conditionally independent tests in the absence of a gold standard.
      was used for computing an approximate sample size for estimating diagnostic accuracy of the PAG ELISA test at 23 to 27 d with a given precision. These computations were done using information from
      • Dufour S.
      • Durocher J.
      • Dubuc J.
      • Dendukuri N.
      • Hassan S.
      • Buczinski S.
      Bayesian estimation of sensitivity and specificity of a milk pregnancy-associated glycoprotein-based ELISA and of transrectal ultrasonographic exam for diagnosis of pregnancy at 28–45 days following breeding in dairy cows.
      for a scenario comparing the PAG test at 23 to 27 d with a PAG test or a US exam (i.e., 2 populations of cows) conducted later. For these calculations, we assumed a Se and Sp of 0.95 and 0.99, respectively, for both the 23 to 27 d PAG test and the 2 confirmation tests. We specified a small difference (<20 percentage points) between the 2 populations' pregnancy prevalence. From these values, it was estimated that a sample size of 191 and 64 cows in each group, so a total of 255 cows, would allow for estimating the tests' Se and Sp with a precision of ± 13 percentage points. For this sample size estimation, however, we did not consider the potential conditional dependence between tests, nor the possibility of pregnancy loss between initial and reference tests (which would likely increase the required sample size). In addition, we did not consider that informative priors would be used for many of the unknown parameters (which would likely decrease the require sample size).

      RESULTS

      A total of 268 samples were collected from 257 cows for the study. There were 247 cows that had only one test conducted, 9 cows were tested for 2 different inseminations, and one cow was tested for 3 different inseminations. The 23 to 27 d PAG ELISA test was conducted, on average (SD) at 25.7 (1.0) d postinsemination (Figure 2). For 200 cows the reference test was a PAG ELISA in milk ≥28 d after insemination, whereas it was an US exam ≥28 d after insemination for 68 cows. The reference tests were conducted, on average (SD) at 33.3 (4.0) d after insemination for cows confirmed with a milk PAG ELISA, and at 37.8 (7.2) d after insemination for cows with a US exam used as reference test (Figure 3). The average (SD) time difference between the PAG ELISA test at 23 to 27 d and the reference test was 7.6 (4.0) d, when the reference test was a PAG ELISA ≥28 d, and 11.8 (7.0) d when the reference test was a US exam.
      Figure thumbnail gr2
      Figure 2Standardized corrected optical density (OD) values of a pregnancy-associated glycoprotein (PAG) ELISA test in milk, conducted between 23 and 27 d after insemination in lactating Holstein dairy cows, as function of the result of a reference test (PAG ELISA for 200 samples and an ultrasound exam for 68 samples) conducted ≥28 d postinsemination. The whole data set (A) is illustrated as well as a subset (B) comprising only samples (n = 198) with OD values ≤0.60 to better illustrate the distribution in the range of OD values that could be used to establish a cut point.
      Figure thumbnail gr3
      Figure 3Distribution of number of days after insemination when the confirmation test was conducted as function of whether the reference test was a pregnancy-associated glycoprotein (PAG) ELISA test or ultrasound (US) exam. Both reference tests were done ≥28 d after insemination.
      The distribution of standardized corrected OD values of the PAG ELISA test at 23 to 27 d as function of the number of days since insemination and of pregnancy status at confirmation ≥28 d postinsemination is illustrated in Figure 2. Most of the apparent disagreements between the PAG ELISA at 23 to 27 d and reference tests appear to be cows with high standardized corrected OD values between 23 and 27 d after insemination, but subsequently confirmed open at ≥28 postinsemination. A few cows, however, had relatively low standardized corrected OD values between 23 and 27 d after insemination, but were later diagnosed pregnant at ≥28 d postinsemination.
      The cross-classified test results, for the potential milk PAG ELISA standardized corrected OD values cut-points at 23 to 27 d after insemination and as function of the reference test used (PAG ELISA ≥28 d or US ≥28 d) are presented in Table 1.
      Table 1Cross-classified test results between a pregnancy-associated glycoprotein (PAG1) ELISA test in milk 23 to 27 d after insemination and either a subsequent similar ELISA test (PAG2; n = 200 samples) or an ultrasound exam (US; 68 samples) conducted between 28 and 59 d postinsemination
      The cross-classified results are presented for PAG1 standardized corrected optical density (OD) values from 0.05 to 0.40 in increments of 0.05. For PAG2, a threshold of OD ≥0.25 was used to classify cows as pregnant (+).
      PAG1 OD thresholdCows confirmed with a PAG ELISA ≥28 d after inseminationCows confirmed with US exam ≥28 d after insemination
      PAG1+ PAG2+PAG1+ PAG2−PAG1− PAG2+PAG1− PAG2−PAG1+ US+PAG1+ US−PAG1− US+PAG1− US−
      0.0510330166479012
      0.1010114382474017
      0.151008488472019
      0.209461090461120
      0.258362190451220
      0.307832693411620
      0.357323194401720
      0.4065139953511220
      1 The cross-classified results are presented for PAG1 standardized corrected optical density (OD) values from 0.05 to 0.40 in increments of 0.05. For PAG2, a threshold of OD ≥0.25 was used to classify cows as pregnant (+).
      All LCM ran with no convergence issues. The smallest effective sample size (3,928) was observed for the Se estimate of the PAG ELISA test 23 to 27 d after insemination using a cutoff of 0.05. For all other cut-points, effective sample sizes were between 6,354 and 60,000 for describing the posterior distributions of the unknown parameters.
      A summary of the Se, Sp, NPV, and PPV estimates obtained with the different PAG ELISA standardized corrected OD thresholds at 23 to 27 d after AI is presented in Figure 4. Briefly, a sharp increase in the Sp of the test was observed between standardized corrected OD of 0.05 and 0.15. Beyond 0.15, the gain in Sp was small. Sensitivity was stable between standardized corrected OD of 0.05 and 0.15 but declined sharply beyond 0.15. The patterns of PPV and NPV were similar because the true prevalence of pregnancy was estimated to be around 0.50. Accordingly, the optimal OD cut point for classification of pregnancy between 23 and 27 d after AI was 0.15. Varying the OD threshold had very little effect on the estimated prevalence of pregnant cows, thus confirming the stability of the LCM. The estimated median prevalence ranged from 0.51 to 0.56, when PAG at 28 to 59 d was the reference test, and from 0.67 to 0.73, when US was the reference test (Supplemental File S1).
      Figure thumbnail gr4
      Figure 4Summary of the median and 95% credible intervals of the sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV) of a pregnancy-associated glycoprotein (PAG) ELISA test conducted between 23 and 27 d after insemination in dairy cows, using different PAG standardized corrected optical density (OD) thresholds for test interpretation. Estimates were obtained using a latent class model considering the imperfection of the reference tests used for comparison, the probability of pregnancy loss between the test at 23 to 27 d and the reference test conducted ≥28 d after insemination, and the conditional dependency between the initial and reference PAG tests. The posterior distributions of the prevalence of pregnancy (median prevalence ranged from 0.51 to 0.56) at 23 to 27 d after insemination in cows confirmed using a subsequent PAG test ≥28 d after insemination were used to compute the NPV and PPV. Thus, NPV and PPV were not fixed attributes of the diagnostic test accuracy. Based on these results, we used a cut point of OD >0.15 to classify cows pregnant based on the PAG ELISA test between 23 and 27 d after insemination.
      Results from the model applying a standardized corrected OD threshold of 0.15 for PAG ELISA test in milk 23 to 27 d after insemination are presented in Table 2. Results from the other investigated thresholds are in Supplemental File S1. Both Se and Sp of the PAG ELISA test in milk 23 to 27 d after AI were excellent, with median (95% CrI) estimates of 0.98 (0.93–1.0) and 0.98 (0.89–1.0), respectively. The true prevalence of pregnancy between 23 to 27 d postinsemination was estimated at 0.54 (0.46–0.61) in cows where the confirmation test was a PAG ELISA test ≥28 d, and at 0.68 (0.58–0.78) in cows where the confirmation test was US ≥28 d.
      Table 2Results from latent class models to determine the diagnostic accuracy of a pregnancy associated glycoprotein (PAG1) ELISA test in milk 23 to 27 d after insemination when used with a threshold of optical density (OD) ≥0.15 to classify cows as pregnant
      The reference test was either a PAG ELISA test 28 to 59 d (mean 33 d) after insemination (PAG2; n = 200 samples) or ultrasound exam 28 to 59 d (mean 38 d) after insemination (n = 68 samples). The informative model was run using informative prior distributions obtained from the literature or from available independent data sets. Two additional models were run as sensitivity analyses: (1) in the model with perturbed priors, the previous priors were modified to change their central value and make them more diffuse; and (2) in the model assuming no pregnancy loss between tests, the probability of pregnancy loss was set to be exactly 0. CrI = 95% credible intervals.
      ParameterInformative modelModel with perturbed priorsModel assuming no pregnancy loss between tests
      PriorEstimatePriorEstimatePriorEstimate
      Median95% CrI
      The reference test was either a PAG ELISA test 28 to 59 d (mean 33 d) after insemination (PAG2; n = 200 samples) or ultrasound exam 28 to 59 d (mean 38 d) after insemination (n = 68 samples). The informative model was run using informative prior distributions obtained from the literature or from available independent data sets. Two additional models were run as sensitivity analyses: (1) in the model with perturbed priors, the previous priors were modified to change their central value and make them more diffuse; and (2) in the model assuming no pregnancy loss between tests, the probability of pregnancy loss was set to be exactly 0. CrI = 95% credible intervals.
      Median95% CrIMedian95% CrI
      Se
      Se = sensitivity.
      PAG1
      PAG1 = PAG ELISA test in milk 23–27 d after insemination.
      Beta (1,1)0.980.93, 1.0Beta (1,1)0.980.91, 1.0Beta (1,1)0.980.93, 1.0
      Sp
      Sp = specificity.
      PAG1
      Beta (1,1)0.980.89, 1.0Beta (1,1)0.940.84, 1.0Beta (1,1)0.910.84, 0.99
      NPV
      Negative (NPV) and positive predictive value (PPV) computed using the estimated prevalence of pregnancy between 23 and 27 d after inseminations in breedings where PAG2 was used as the reference test.
      PAG1
      NA
      NA = not applicable.
      0.980.91, 1.0NA0.970.88, 1.0NA0.980.92, 1.0
      PPV
      Negative (NPV) and positive predictive value (PPV) computed using the estimated prevalence of pregnancy between 23 and 27 d after inseminations in breedings where PAG2 was used as the reference test.
      PAG1
      NA0.980.90, 1.0NA0.950.85, 1.0NA0.920.85, 0.99
      Se PAG2
      PAG2 = PAG ELISA test in milk ≥28 d after insemination.
      Beta (100, 2)0.980.95, 1.0Beta (43, 6)
      Asterisks indicate main differences between the informative model and the models used in the sensitivity analyses.
      0.910.84, 0.96Beta (100, 2)0.980.94, 1.0
      Sp PAG2Beta (92, 5)0.950.91, 0.98Beta (46, 8)
      Asterisks indicate main differences between the informative model and the models used in the sensitivity analyses.
      0.810.77, 0.93Beta (92, 5)0.950.91, 0.98
      Se US
      US ultrasound exam ≥28 d after insemination.
      Beta (100, 7)0.950.91, 0.98Beta (47, 10)
      Asterisks indicate main differences between the informative model and the models used in the sensitivity analyses.
      0.900.83, 0.95Beta (100, 7)0.950.90, 0.98
      Sp USBeta (100, 6)0.950.90, 0.98Beta (46, 9)
      Asterisks indicate main differences between the informative model and the models used in the sensitivity analyses.
      0.880.79, 0.94Beta (100, 6)0.950.91, 0.98
      P23–27 PAG2
      P23–27 PAG2 = true prevalence of pregnancy between 23 and 27 d post insemination in breedings where PAG2 was used as the reference test.
      Beta (9, 8)0.540.46, 0.61Beta (9, 8)0.530.44, 0.61Beta (9, 8)0.510.44, 0.58
      P23–27 US
      P23–27 US = true prevalence of pregnancy between 23 and 27 d post insemination in breedings where US was used as the reference test.
      Beta (9, 8)0.680.58, 0.78Beta (9, 8)0.680.58, 0.78Beta (9, 8)0.680.57, 0.77
      PLoss
      PLoss = probability of pregnancy loss between the PAG1 test and subsequent PAG2 or US tests.
      Beta (6, 43)0.070.03, 0.11Beta (6, 43)0.060.02, 0.12NA
      Asterisks indicate main differences between the informative model and the models used in the sensitivity analyses.
      NANA
      Covp
      Covariance between Se (covp) and between Sp (covn) of PAG1 and PAG2.
      Unif (a, b)
      Where a is (1 − PAG1 Se) × (Se PAG2 − 1); b is the smallest Se between PAG1 and PAG2 minus the product of these Se; c is (PAG1 Sp − 1) × (1 − PAG2 Sp); and d is the smallest Sp between PAG1 and PAG2 minus the product of these Sp.
      0.000.00, 0.02Unif (a, b)0.010.00, 0.06Unif (a, b)0.000.00, 0.02
      Covn
      Covariance between Se (covp) and between Sp (covn) of PAG1 and PAG2.
      Unif (c, d)
      Where a is (1 − PAG1 Se) × (Se PAG2 − 1); b is the smallest Se between PAG1 and PAG2 minus the product of these Se; c is (PAG1 Sp − 1) × (1 − PAG2 Sp); and d is the smallest Sp between PAG1 and PAG2 minus the product of these Sp.
      0.010.00, 0.03Unif (c, d)0.030.00, 0.08Unif (c, d)0.010.00, 0.03
      1 The reference test was either a PAG ELISA test 28 to 59 d (mean 33 d) after insemination (PAG2; n = 200 samples) or ultrasound exam 28 to 59 d (mean 38 d) after insemination (n = 68 samples). The informative model was run using informative prior distributions obtained from the literature or from available independent data sets. Two additional models were run as sensitivity analyses: (1) in the model with perturbed priors, the previous priors were modified to change their central value and make them more diffuse; and (2) in the model assuming no pregnancy loss between tests, the probability of pregnancy loss was set to be exactly 0. CrI = 95% credible intervals.
      2 Se = sensitivity.
      3 PAG1 = PAG ELISA test in milk 23–27 d after insemination.
      4 Sp = specificity.
      5 Negative (NPV) and positive predictive value (PPV) computed using the estimated prevalence of pregnancy between 23 and 27 d after inseminations in breedings where PAG2 was used as the reference test.
      6 NA = not applicable.
      7 PAG2 = PAG ELISA test in milk ≥28 d after insemination.
      8 US ultrasound exam ≥28 d after insemination.
      9 P23–27 PAG2 = true prevalence of pregnancy between 23 and 27 d post insemination in breedings where PAG2 was used as the reference test.
      10 P23–27 US = true prevalence of pregnancy between 23 and 27 d post insemination in breedings where US was used as the reference test.
      11 PLoss = probability of pregnancy loss between the PAG1 test and subsequent PAG2 or US tests.
      12 Covariance between Se (covp) and between Sp (covn) of PAG1 and PAG2.
      13 Where a is (1 − PAG1 Se) × (Se PAG2 − 1); b is the smallest Se between PAG1 and PAG2 minus the product of these Se; c is (PAG1 Sp − 1) × (1 − PAG2 Sp); and d is the smallest Sp between PAG1 and PAG2 minus the product of these Sp.
      * Asterisks indicate main differences between the informative model and the models used in the sensitivity analyses.
      The results were not sensitive to the choice of prior used for estimating the diagnostic accuracy of the PAG ELISA test 23 to 27 d after insemination. In most cases the median Se and Sp estimates changed by 1 or 2 percentage points (Table 2 and Supplemental File S3). In the model with perturbed priors and using an OD threshold of 0.15, the median Sp estimate was reduced by 4 percentage points (0.94, 95% CrI: 0.84–1.0; Table 2).
      Assuming that there was no change in pregnancy status between the test at 23 to 27 d and the reference test ≥28 d after insemination led to underestimation of both the Sp of the test (0.91; 95% CrI: 0.84, 0.99) and of the prevalence of pregnancy at 23 to 27 d after AI (Table 2 and Supplemental File S4).

      DISCUSSION

      Our main finding is that a commercially available PAG ELISA in milk may be a viable option for pregnancy diagnosis between 23 and 27 d after insemination. If confirmed with additional data, this could save up to 5 d over the current typical practice, which is pregnancy diagnosis by US or measurement of PAG from 28 d after insemination. We underline that the utility of earlier identification of nonpregnancy will depend on herd-specific logistics, turn-around time from sample collection to receipt of results, and selected management tactics for timely and fertile re-insemination of nonpregnant cows. The main limitation of our study was the unavailability of a gold-standard test to which we could directly compare the test under investigation. However, valid methods exist to compare and estimate the test characteristics of multiple imperfect tests (
      • Cheung A.
      • Dufour S.
      • Jones G.
      • Kostoulas P.
      • Stevenson M.A.
      • Singanallur N.B.
      • Firestone S.M.
      Bayesian latent class analysis when the reference test is imperfect.
      ). Moreover, we were able to implement a generalization of the classical LCM in which the probability of a change in status of the animal between tests (pregnancy loss in our case) is also explicitly considered. We observed that ignoring this potential change in pregnancy status would have led to underestimation of the specificity of PAG ELISA test at 23 to 27 d after insemination. Furthermore, the informative priors used in our models were informed by very rigorous sources and we were able to demonstrate that our estimates were not very sensitive to the choice of prior used. Finally, with the available sample size and because of the relatively strong agreement between tests, we were able to calculate very precise estimates of accuracy.
      We stress that we had only 5 pregnancy tests conducted at exactly 23 d after insemination. Although the pattern of results appeared consistent between d 23 and 27 (Figure 2), the data were sparse before d 25. As discussed below, the biology of PAG in cattle supports the possibility of pregnancy diagnosis as early as d 22, but there is more robust evidence ≥25 d after insemination. We encourage additional validation studies of PAG ELISA tests between 22 and 28 d after AI before widespread application of these tests at that time. Nevertheless, the very high Se and Sp in our results are encouraging for this application.
      The true prevalence of pregnancy between 23 and 27 d after insemination was estimated at 0.54 (95% CrI: 0.46, 0.61) and 0.68 (95% CrI: 0.58, 0.78) when the confirmation test was PAG ELISA ≥28 d or US ≥28 d, respectively. The apparent difference is likely an artifact in the data because one of participating farmers changed his management practice during the study from using PAG ELISA ≥28 d as soon as possible on all cows tested between 23 and 27 d after insemination to waiting for the monthly US examinations for cows he felt had high PAG level at the test done 23 to 27 d after insemination. Cows with lower PAG values at 23 to 27 d after insemination were tested with PAG ELISA ≥28 as soon as possible to allow for rapid rebreeding if confirmed open.
      In studies with repeated sampling, PAG values rose rapidly between 24 and 28 d after AI in pregnant cows (
      • Green J.A.
      • Parks T.E.
      • Avalle M.P.
      • Telugu B.P.
      • McLain A.L.
      • Peterson A.J.
      • McMillan W.
      • Mathialagan N.
      • Hook R.R.
      • Xie S.
      • Roberts R.M.
      The establishment of an ELISA for the detection of pregnancy-associated glycoproteins (PAGs) in the serum of pregnant cows and heifers.
      ;
      • Friedrich M.
      • Holtz W.
      Establishment of an ELISA for measuring bovine pregnancy-associated glycoprotein in serum or milk and its application for early pregnancy detection.
      ) leading gradually to significant differences between pregnant and nonpregnant cows. The NPV of a serum PAG ELISA test reported by
      • Friedrich M.
      • Holtz W.
      Establishment of an ELISA for measuring bovine pregnancy-associated glycoprotein in serum or milk and its application for early pregnancy detection.
      were similar between 26 and 30 d (0.93) and 31 to 35 d after AI (0.96) using a cut point of 1.5 ng/mL. When the threshold was changed to 2.0 ng/mL, the NPV of the test decreased meaningfully at 26 to 30 d post-AI (0.77) with little change between 31 and 35 d post-AI (0.93). Data from a previous study (
      • Dufour S.
      • Durocher J.
      • Dubuc J.
      • Dendukuri N.
      • Hassan S.
      • Buczinski S.
      Bayesian estimation of sensitivity and specificity of a milk pregnancy-associated glycoprotein-based ELISA and of transrectal ultrasonographic exam for diagnosis of pregnancy at 28–45 days following breeding in dairy cows.
      ) suggested that a standardized corrected OD value of 0.25 was the optimal cutoff for the current milk PAG-based test under investigation when performed between 28 and 45 d after insemination. The same conclusion was reached by
      • Commun L.
      • Velek K.
      • Barbry J.-B.
      • Pun S.
      • Rice A.
      • Mestek A.
      • Egli C.
      • Leterme S.
      Detection of pregnancy-associated glycoproteins in milk and blood as a test for early pregnancy in dairy cows.
      when cows were sampled between 30 and 41 d after insemination. Data from our study showed Se and Sp (95% CrI) of 0.98 (0.96, 1.00) and 0.98 (0.91, 1 0.00), respectively, when a cutoff standardized corrected OD value of 0.15 was applied between 23 and 27 d postinsemination, compared with 0.91 (0.84, 0.98) and 0.98 (0.93, 1.00) when 0.25 was used as a cutoff. These findings support the hypothesis that the interpretive threshold should be adjusted when performing a PAG ELISA <28 d after insemination. The predictive value of PAG ELISA tests for pregnancy diagnosis or confirmation might be improved by adjusting the cutoff value as function of the number of days since insemination. To validate this hypothesis, however, studies with a large number of cows sampled at various numbers of d postinsemination will be needed.
      A novel approach based on within-cow variation of serum pregnancy-specific protein B (PSPB) compared with basal level was recently tested to diagnose pregnancy before 28 d postinsemination.
      • Martins J.P.N.
      • Wang D.
      • Mu N.
      • Rossi G.F.
      • Martini A.P.
      • Martins V.R.
      • Pursley J.R.
      Level of circulating concentrations of progesterone during ovulatory follicle development affects timing of pregnancy loss in lactating dairy cows.
      showed that an increase >28% at d 23 above a basal level of the mean at d 16 and 20 was a better indicator of pregnancy status (Se = 0.98; Sp = 0.97) than a single PSPB test on d 23 (Se = 0.93; Sp = 0.97).
      • Middleton E.L.
      • Pursley J.R.
      Short communication: Blood samples before and after embryonic attachment accurately determine non-pregnant lactating dairy cows at 24 d post-artificial insemination using a commercially available assay for pregnancy-specific protein B.
      simplified the within-cow variation approach with a single measurement to establish the basal serum PSPB level. Cows were classified pregnant at 24 d after insemination when there was >10% increase in serum PSPB between 17 and 24 d after AI, with Se = 1.00 and Sp = 0.94. Despite the high accuracy of this approach, the management and costs associated with this precisely paired sampling might not be practical for some dairy operations.
      Early pregnancy diagnosis must provide accurate identification of nonpregnant cows. Misclassification costs (costs of false negative and false positive diagnoses) are mainly driven by the risk of causing abortion in cows misclassified nonpregnant and of additional days open and the increased risk of culling in cows misclassified pregnant. Precise calculation of these costs is complex and depends on numerous variables including the stages of lactation and of pregnancy, but generally a false negative pregnancy diagnosis is more detrimental than false positive because they will often lead to prostaglandin injection and iatrogenic pregnancy loss. Breeding a pregnant cow is also a risk factor for pregnancy loss (
      • Moore D.A.
      • Overton M.W.
      • Chebel R.C.
      • Truscott M.L.
      • BonDurant R.H.
      Evaluation of factors that affect embryonic loss in dairy cattle.
      ). The effect of false positive diagnosis on days open is mitigated in most dairy operations by running a confirmation test around 60 d after insemination.
      • Giordano J.O.
      • Fricke P.M.
      • Cabrera V.E.
      Economics of resynchronization strategies including chemical tests to identify nonpregnant cows.
      showed that the Se of a chemical test used for early pregnancy diagnosis would have a greater effect on its economic value than Sp. The objective when performing early pregnancy diagnosis should be to maximize the negative predictive value (correctly identify open cows) without compromising the PPV, and maximizing Se without compromising Sp will help achieve this. Although Se and Sp of the milk PAG ELISA test under investigation between 23 and 27 d after insemination were both excellent (98%), we encourage replication of this work with a larger sample.

      CONCLUSIONS

      This study showed the potential of a commercial milk PAG-based ELISA to accurately diagnose pregnancy in dairy cows 23 to 27 d following insemination. Based on the present analyses, a cutoff corresponding to an OD of 0.15 can be suggested. We advocate additional validation studies on larger samples to confirm the optimal cut point(s) and the test performance before implementation of the test between 23 and 27 d after insemination and to evaluate the conditions under which earlier testing is beneficial.

      ACKNOWLEDGMENTS

      This project was financially supported in part by IDEXX Laboratories (Westbrook, ME). The senior author (SD) was supported by a Discovery grant from the Natural Sciences and Engineering Research Council of Canada (RGPIN-2020-05237; Ottawa, Canada). We thank the participating producers for their involvement in the sampling and shipping procedures, and the use of their data. We also thank the laboratory staff from Lactanet (Québec, Canada) for their dedicated work. The authors have not stated any conflicts of interest.

      REFERENCES

        • Cerri R.L.A.
        • Juchem S.O.
        • Chebel R.C.
        • Rutigliano H.M.
        • Bruno R.G.S.
        • Galvão K.N.
        • Thatcher W.W.
        • Santos J.E.P.
        Effect of fat source differing in fatty acid profile on metabolic parameters, fertilization, and embryo quality in high-producing dairy cows.
        J. Dairy Sci. 2009; 92: 1520-1531
        • Chebel R.C.
        • Santos J.E.P.
        • Cerri R.L.A.
        • Galvão K.N.
        • Juchem S.O.
        • Thatcher W.W.
        Effect of resynchronization with GnRH on day 21 after artificial insemination on pregnancy rate and pregnancy loss in lactating dairy cows.
        Theriogenology. 2003; 60 (14519461): 1389-1399
        • Cheung A.
        • Dufour S.
        • Jones G.
        • Kostoulas P.
        • Stevenson M.A.
        • Singanallur N.B.
        • Firestone S.M.
        Bayesian latent class analysis when the reference test is imperfect.
        Rev. Sci. Tech. 2021; 40 (34140724): 271-286
        • Commun L.
        • Velek K.
        • Barbry J.-B.
        • Pun S.
        • Rice A.
        • Mestek A.
        • Egli C.
        • Leterme S.
        Detection of pregnancy-associated glycoproteins in milk and blood as a test for early pregnancy in dairy cows.
        J. Vet. Diagn. Invest. 2016; 28 (26944877): 207-213
        • Dendukuri N.
        • Joseph L.
        Bayesian approaches to modeling the conditional dependence between multiple diagnostic tests.
        Biometrics. 2001; 57 (11252592): 158-167
        • Dufour S.
        • Durocher J.
        • Dubuc J.
        • Dendukuri N.
        • Hassan S.
        • Buczinski S.
        Bayesian estimation of sensitivity and specificity of a milk pregnancy-associated glycoprotein-based ELISA and of transrectal ultrasonographic exam for diagnosis of pregnancy at 28–45 days following breeding in dairy cows.
        Prev. Vet. Med. 2017; 140 (28460745): 122-133
        • Durocher J.
        • Moore R.K.
        • Albaaj A.
        • Leblanc S.
        • Dufour S.
        Bayesian estimation of sensitivity and specificity of a commercial milk pregnancy-associated glycoprotein ELISA test between 23 and 27 days after breeding in dairy cows.
        Scholars Portal Dataverse. 2021; V1
        • Fosgate G.T.
        • Motimele B.
        • Ganswindt A.
        • Irons P.C.
        A Bayesian latent class model to estimate the accuracy of pregnancy diagnosis by transrectal ultrasonography and laboratory detection of pregnancy-associated glycoproteins in dairy cows.
        Prev. Vet. Med. 2017; 145 (28903866): 100-109
        • Fricke P.M.
        • Ricci A.
        • Giordano J.O.
        • Carvalho P.D.
        Methods for and implementation of pregnancy diagnosis in dairy cows.
        Vet. Clin. North Am. Food Anim. Pract. 2016; 32 (26922118): 165-180
        • Friedrich M.
        • Holtz W.
        Establishment of an ELISA for measuring bovine pregnancy-associated glycoprotein in serum or milk and its application for early pregnancy detection.
        Reprod. Domest. Anim. 2010; 45 (19032429): 142-146
        • Galvão K.N.
        • Santos J.E.P.
        • Juchem S.O.
        • Cerri R.L.A.
        • Coscioni A.C.
        • Villaseñor M.
        Effect of addition of a progesterone intravaginal insert to a timed insemination protocol using estradiol cypionate on ovulation rate, pregnancy rate, and late embryonic loss in lactating dairy cows.
        J. Anim. Sci. 2004; 82 (15537771): 3508-3517
        • Georgiadis M.P.
        • Johnson W.O.
        • Gardner I.A.
        Sample size determination for estimation of the accuracy of two conditionally independent tests in the absence of a gold standard.
        Prev. Vet. Med. 2005; 71 (16076507): 1-10
        • Giordano J.O.
        • Fricke P.M.
        • Cabrera V.E.
        Economics of resynchronization strategies including chemical tests to identify nonpregnant cows.
        J. Dairy Sci. 2013; 96 (23261387): 949-961
        • Green J.A.
        • Parks T.E.
        • Avalle M.P.
        • Telugu B.P.
        • McLain A.L.
        • Peterson A.J.
        • McMillan W.
        • Mathialagan N.
        • Hook R.R.
        • Xie S.
        • Roberts R.M.
        The establishment of an ELISA for the detection of pregnancy-associated glycoproteins (PAGs) in the serum of pregnant cows and heifers.
        Theriogenology. 2005; 63 (15725453): 1481-1503
        • Green J.A.
        • Xie S.
        • Roberts R.M.
        Pepsin-related molecules secreted by trophoblast.
        Rev. Reprod. 1998; 3 (9509990): 62-69
        • Johnson W.O.
        • Jones G.
        • Gardner I.A.
        Gold standards are out and Bayes is in: Implementing the cure for imperfect reference tests in diagnostic accuracy studies.
        Prev. Vet. Med. 2019; 167 (31027713): 113-127
        • Karen A.
        • Sousa N.M.D.
        • Beckers J.-F.
        • Bajcsy Á.C.
        • Tibold J.
        • Mádl I.
        • Szenci O.
        Comparison of a commercial bovine pregnancy-associated glycoprotein ELISA test and a pregnancy-associated glycoprotein radioimmunoassay test for early pregnancy diagnosis in dairy cattle.
        Anim. Reprod. Sci. 2015; 159 (26059776): 31-37
        • Kostoulas P.
        • Nielsen S.S.
        • Branscum A.J.
        • Johnson W.O.
        • Dendukuri N.
        • Dhand N.K.
        • Toft N.
        • Gardner I.A.
        STARD-BLCM: Standards for the Reporting of Diagnostic accuracy studies that use Bayesian Latent Class Models.
        Prev. Vet. Med. 2017; 138 (28237234): 37-47
        • Lawson B.C.
        • Shahzad A.H.
        • Dolecheck K.A.
        • Martel E.L.
        • Velek K.A.
        • Ray D.L.
        • Lawrence J.C.
        • Silvia W.J.
        A pregnancy detection assay using milk samples: Evaluation and considerations.
        J. Dairy Sci. 2014; 97 (25129499): 6316-6325
        • Martins J.P.N.
        • Wang D.
        • Mu N.
        • Rossi G.F.
        • Martini A.P.
        • Martins V.R.
        • Pursley J.R.
        Level of circulating concentrations of progesterone during ovulatory follicle development affects timing of pregnancy loss in lactating dairy cows.
        J. Dairy Sci. 2018; 101 (30197145): 10505-10525
        • Mercadante P.M.
        • Ribeiro E.S.
        • Risco C.
        • Ealy A.D.
        Associations between pregnancy-associated glycoproteins and pregnancy outcomes, milk yield, parity, and clinical diseases in high-producing dairy cows.
        J. Dairy Sci. 2016; 99 (26851856): 3031-3040
        • Middleton E.L.
        • Pursley J.R.
        Short communication: Blood samples before and after embryonic attachment accurately determine non-pregnant lactating dairy cows at 24 d post-artificial insemination using a commercially available assay for pregnancy-specific protein B.
        J. Dairy Sci. 2019; 102 (31178191): 7570-7575
        • Moore D.A.
        • Overton M.W.
        • Chebel R.C.
        • Truscott M.L.
        • BonDurant R.H.
        Evaluation of factors that affect embryonic loss in dairy cattle.
        J. Am. Vet. Med. Assoc. 2005; 226 (15825738): 1112-1118
        • Moreira F.
        • Orlandi C.
        • Risco C.A.
        • Mattos R.
        • Lopes F.
        • Thatcher W.W.
        Effects of presynchronization and bovine somatotropin on pregnancy rates to a timed artificial insemination protocol in lactating dairy cows.
        J. Dairy Sci. 2001; 84 (11467815): 1646-1659
        • Oliveira Filho R.V.
        • Franco G.A.
        • Reese S.T.
        • Dantas F.G.
        • Fontes P.L.P.
        • Cooke R.F.
        • Rhinehart J.D.
        • Thompson K.W.
        • Pohler K.G.
        Using pregnancy associated glycoproteins (PAG) for pregnancy detection at day 24 of gestation in beef cattle.
        Theriogenology. 2020; 141 (31539641): 128-133
        • Reese S.T.
        • Pereira M.H.C.
        • Edwards J.L.
        • Vasconcelos J.L.M.
        • Pohler K.G.
        Pregnancy diagnosis in cattle using pregnancy associated glycoprotein concentration in circulation at day 24 of gestation.
        Theriogenology. 2018; 106 (29073542): 178-185
        • Ricci A.
        • Carvalho P.D.
        • Amundson M.C.
        • Fourdraine R.H.
        • Vincenti L.
        • Fricke P.M.
        Factors associated with pregnancy-associated glycoprotein (PAG) levels in plasma and milk of Holstein cows during early pregnancy and their effect on the accuracy of pregnancy diagnosis.
        J. Dairy Sci. 2015; 98 (25660740): 2502-2514
        • Romano J.E.
        • Thompson J.A.
        • Forrest D.W.
        • Westhusin M.E.
        • Tomaszweski M.A.
        • Kraemer D.C.
        Early pregnancy diagnosis by transrectal ultrasonography in dairy cattle.
        Theriogenology. 2006; 66 (16584765): 1034-1041
        • Santos J.E.
        • Thatcher W.W.
        • Pool L.
        • Overton M.W.
        Effect of human chorionic gonadotropin on luteal function and reproductive performance of high-producing lactating Holstein dairy cows.
        J. Anim. Sci. 2001; 79 (11768118): 2881-2894
        • Santos J.E.P.
        • Bartolome J.A.
        • Cerri R.L.A.
        • Juchem S.O.
        • Hernandez O.
        • Trigg T.
        • Thatcher W.W.
        Effect of a deslorelin implant in a timed artificial insemination protocol on follicle development, luteal function and reproductive performance of lactating dairy cows.
        Theriogenology. 2004; 61 (14662141): 421-435
        • Santos J.E.P.
        • Thatcher W.W.
        • Chebel R.C.
        • Cerri R.L.A.
        • Galvão K.N.
        The effect of embryonic death rates in cattle on the efficacy of estrus synchronization programs.
        Anim. Reprod. Sci. 2004; 82–83 (15271477): 513-535
        • Silke V.
        • Diskin M.G.
        • Kenny D.A.
        • Boland M.P.
        • Dillon P.
        • Mee J.F.
        • Sreenan J.M.
        Extent, pattern and factors associated with late embryonic loss in dairy cows.
        Anim. Reprod. Sci. 2002; 71 (11988367): 1-12
        • Sturtz S.
        • Ligges U.
        • Gelman A.
        R2WinBUGS: A package for running WinBUGS from R.
        J. Stat. Softw. 2005; 12
        • Vasconcelos J.L.M.
        • Silcox R.W.
        • Lacerda J.A.
        • Pursley J.R.
        • Wiltbank M.C.
        Pregnancy rate, pregnancy loss, and response to head stress after AI at 2 different times from ovulation in dairy cows.
        Biol. Reprod. 1997; 56: 230
        • Wiltbank M.C.
        • Baez G.M.
        • Garcia-Guerra A.
        • Toledo M.Z.
        • Monteiro P.L.J.
        • Melo L.F.
        • Ochoa J.C.
        • Santos J.E.P.
        • Sartori R.
        Pivotal periods for pregnancy loss during the first trimester of gestation in lactating dairy cows.
        Theriogenology. 2016; 86 (27238438): 239-253