Reproduction in Domestic Animals

118 NC Friggens, M Bjerring, C Ridder, S Højsgaard and T Larsenbasic specificity of 98.7%. However, this assumes thatfor any given progesterone observation oestrus could beindicated. This was not the case, because in the model,once oestrus was indicated, a waiting period of 5 dayswas mandatory before a new oestrus can be indicated(Friggens and Chagunda 2005). Allowing for thiswaiting period, the total number of progesterone measurementsfor which oestrus could be determined was2 723 resulting in a specificity of 93.7%. This specificityis similar to that found by de Mol et al. (1999).Examination of the false positives showed that theywere largely associated with low progesterone concentrationsfluctuating around the 4 ng ⁄ ml threshold(Fig. 3). Nearly 53% (90 ⁄ 171) of the false positivescame at the end of ‘oestrus’ cycles in which themaximum progesterone concentration did not exceed10 ng ⁄ ml. Only in 56 out of the 171 cases had themaximum progesterone concentration in the preceedingcycle exceeded 15 ng ⁄ ml. Clearly, every time such afluctuating profile decreased below 4 ng ⁄ ml, the modelindicated oestrus. This is the major drawback of asimple threshold-based rule, which cannot be overcomeby a simple smoothing of the progesterone profiles. Thenumber of false positives could be reduced by simplyusing a much greater threshold, thus increasing modelspecificity. Unfortunately, a greater threshold has asignificant unwanted cost in terms of reducing modelsensitivity (de Mol et al. 2001). Instead, each timeoestrus is detected the model calculates a likelihood of apotential insemination succeeding. Because this calculationincludes information about the height and length ofthe preceding oestrus cycle, these false positive oestrusesare associated with a very low likelihood of inseminationsucceeding. Average values of the predicted likelihoodof a potential insemination succeeding when the maximumprogesterone concentration in the preceding cycledid not exceed 5, 10 or 15 ng ⁄ ml were: 0.01, 0.17 and0.45 respectively (on a scale 0–0.9). Thus, false positivesProgesterone (ng/ml)30241812600 30 60 90 120 150 180Days from calvingFig. 3. An example of a progesterone profile showing a fluctuating lowprogesterone after calving. Because the smoothed progesterone profile(solid line) repeatedly decreases below 4 ng/ml during this period, themodel repeatedly detected false positive oestruses. Raw, unsmoothedprogesterone values are shown by the open circles. Pluses indicatemodel-determined Status 0 (postpartum an oestrus), solid trianglesindicate Status 1 (oestrus cycling), and solid squares indicate Status 2(potentially pregnant). The downward pointing open triangles indicateinseminations, the upwards pointing open triangles indicate pregnancydeterminations (1 ‘‘ng/ml’’ = not pregnant, 9 ‘‘ng/ml’’ = pregnant)are obvious to the end user of the system. The predictedlikelihood of a potential insemination succeeding alsoprovides the end user with helpful information formaking a decision about whether or not to inseminatethe cow. This has been explored further by Friggens andLøvendahl (2008).In the model, the first increase in the smoothedprogesterone profile above 4 ng ⁄ ml indicates the end ofthe postpartum anoestrus and coincidentally the firstoestrus. For those first oestruses that were ratified, theaverage difference between the time of model detectedfirst oestrus and ratified oestrus was 1.5 days(SD = 1.6). The distribution of the differences was leftskewed by those cases in which a period of fluctuationoccurred around the 4-ng ⁄ ml threshold (Fig. 3). Themedian difference was 2.2 days. Thus, the model is ingood agreement with the ratified oestruses. However,because the ‘Mahalanobis template’ was based onoestruses that occurred after a prior period of lutealactivity, it did not identify all first increases in progesteroneabove the threshold. The distribution of thesemodel-detected durations of postpartum anoestrus wassimilar to that reported by Royal et al. (2002).Pregnancy determinationOf the 121 confirmed pregnancies, 108 remained inStatus 2 (model-detected pregnant) from conceptionuntil at least positive pregnancy determination (by rectalpalpation), the sensitivity of the model for detectingpregnancy was thus 89.3%. Of the 11 cases in which themodel falsely detected the cow as no longer pregnant, 2were due to gaps in the progesterone data, 5 failedbecause the model judged that the insemination wasmistimed, and 6 were caused by decreases in progesteroneconcentration. These decreases were in some casesindistinguishable from those one would detect if the cowwas returning to oestrus following early embryo loss(e.g. a sudden decrease to < 10 ng ⁄ ml) but these cowsremained pregnant and subsequently calved. Giventhese types of cases, it is hard to envisage being ableto substantially improve pregnancy detection withoutsimultaneously increasing the number of false negatives.However, of greater important to the end user isdetecting true pregnancy failure.The distribution of model detected pregnancy lengths,for those pregnancies that did not result in a positivepregnancy determination or a calf, provides usefulinformation (Fig. 4). The distribution of time frominsemination to model-detected pregnancy failure isbimodal reflecting the two reasons for the modelassessing the cow to not be pregnant: (1) inappropriatetiming of insemination and (2) progesterone concentrationstoo low (first and second peaks, respectively). Themedian of the distribution was 22 days, reflecting thesecond peak. By 27 days 60% of pregnancy failures hadoccurred, and by 34 days 70% had occurred. Thesefigures reflect a typical time pattern of early embryonicloss (Sreenan et al. 2001). Reliable detection of pregnancyusing progesterone is not possible much earlierthan 18–23 days post-insemination because the progesteroneprofile during the luteal phase is not affected bypresence of an embryo (Bulman and Lamming 1978).Ó 2008 The Authors. Journal compilation Ó 2008 Blackwell Verlag