x
ISSN 0936-6768
was analysed using the Ridgeway ELISA-kit (Ridgeway are also included as inputs. To make use of other known
Science Ltd, Gloucestershire, UK). Milk samples were effectors of reproductive performance that are not
pipetted, diluted and distributed using a Biomek 2000 necessarily reflected in progesterone concentrations,
(Laboratory Automation Workstation, Beckman Coul- the model is designed to incorporate a number of
ter, Fullerton, CA, USA). Milk samples (25 ll, diluted additional inputs. However, the model is designed to
1 + 2 with water) were handled according to the function in the absence of these additional inputs
manufacturer’s instructions, however incubation R4 (although with some loss of accuracy in some outputs).
was increased to 1 h 30 min. Plates were read using a The model outputs are all reproductive Status specific
spectrophotometer ⁄ fluorometer, Fluostar (BMG Lab- with the exception of days to next sample (DNS) which
technologies, Offenburg, Germany) (575 nm). Analyses is calculated in each model run regardless of reproduc-
were performed in 96-well plates; two sets of seven tive Status. Days to next sample is an important output
standards (0–30 ng ⁄ ml), locally made using milk from with respect to the progesterone measurements being
an ovariectomized cow and ethanolic progesterone automated. It is designed to feedback to the sampling
solutions, and two sets of two control samples were system so that the frequency of milk sampling (i.e.
used for every analysis and plate. For the low and high progesterone measurement) can be varied according to
controls, the intra assay precision (CV%) was 14.9 and the predicted likelihood of a future reproductive event
1.4, respectively; the inter-assay precision (CV%) was such as onset of oestrus cycles. The other model outputs
32.7 and 20.1, respectively; and the average inaccuracy are: risk of prolonged postpartum anoestrus, risk and
(bias) was +0.82 and )0.60 ng ⁄ ml, respectively. type of ovarian cyst (i.e. prolonged luteal or follicular
The time series of milk progesterone measurements phase), onset of oestrus, likelihood of a potential
was examined for gaps longer than 14 days with no insemination succeeding and likelihood of being preg-
progesterone values. All data following such gaps were nant (following oestrus).
excluded. Further, the entire cow-lactation was excluded The shift from Status 0 (postpartum anoestrus) to
if there was insufficient data to identify the end of the Status 1 (oestrus cycling) is caused by two consecutive
postpartum anoestrus period (i.e. the time series did not smoothed progesterone measurements above the thresh-
start at calving and there were less than five measure- old value (LThresh) which is 4 ng ⁄ ml. (If EOD inputs
ments before the cow was detected to have elevated are available these can also provoke this shift). As
progesterone and resumed oestrous cycles). The result- described by Friggens and Chagunda (2005), once a cow
ing data represented information during 578 lactations is in Status 1 the shift to Status 2 is caused by the
from 380 cows. The total number of milk progesterone smoothed progesterone value decreasing below
measurements was 55 036. LThresh. This shift generates a model-detected indica-
Timing of inseminations was based on visual tion of oestrus. If no insemination is recorded in the
detection of oestrus backed up by activity meter following 5 days, the cow automatically reverts to Status
indications (DeLaval, Tumba, Sweden). Farm staff 1. If there is a timely insemination and subsequent
had no access to the progesterone data. Visual progesterone concentrations increase, then the cow
observations for oestrus signs were made daily at remains in status 2 (potentially pregnant) as long as
06:00, 12:00 and 16:00, each period lasted 20 min. Cows the model predicted likelihood of pregnancy remains
showing oestrus earlier than 35 days from calving were above a pregnancy threshold. Typically, if progesterone
not inseminated to that oestrus. All cows were concentrations decrease, then the model shifts back
inseminated to oestrus regardless of other management from Status 2 to Status 1.
factors such as age, milk yield, and whether the cow was The model tested in the present study contained the
designated for subsequent culling. Inseminations were following major modification relative to the model
carried out the same day for oestruses detected at the described by Friggens and Chagunda (2005). Visual
morning observation, and the next morning for inspection of the earliest progesterone profiles collected
oestruses detected in the afternoon and evening. showed a number of oestrus-like decreases in proges-
terone that did not decrease below LThresh. To deal
with this an additional threshold, HThresh was intro-
The model being tested duced (6 ng ⁄ ml). This allowed identification of these
The model being tested here has been described in detail ‘high-progesterone oestruses’ conditional on the maxi-
by Friggens and Chagunda (2005). Briefly, the model is mum progesterone concentration in the preceding cycle
based on the cow always being in one of three being greater than 15 ng ⁄ ml, and the present decrease
reproductive states (Status), these are: postpartum being >10 days since a preceding oestrus.
anoestrus (Status 0), oestrus cycling (Status 1), and The model tested in the present study also contained
potentially pregnant (Status 2). In the model these the following minor modifications relative to the model
statuses are mutually exclusive. The Status the cow was described by Friggens and Chagunda (2005). The slope
found to be in by the previous run of the model is always of the progesterone profile is used, for example in the
the Status at the start of the current run. By default, the function calculating the predicted likelihood of AI
cow is assumed to be in Status 0 at calving. The model is success. The slope can have extreme values particularly
dynamic and deterministic, designed to run each time a with short intervals between measurements, thus an
new trigger input occurs using both the current and adjusted slope (a sigmoid transformation) was used
previous values. The main model inputs are a smoothed instead. Over the normal range this makes virtually no
(extended Kalman filter) progesterone concentration difference but it caps the extremes. In Status 2, when the
and slope. Inseminations and pregnancy determinations calculated likelihood of pregnancy (LikePreg) decreases
below a pregnancy threshold, the cow is deemed no not calved by the end of the experiment period. When
longer pregnant and reverts to Status 1. In the current more than one insemination occurred within a time
implementation of the model, a modification was made window, then the last one was chosen. For the purpose
such that once the cow was more than 30 days after of testing the progesterone model, it was required that
insemination, 2 consecutive low LikePreg (i.e. low confirmed oestruses had more than 2 progesterone
progesterone) values are required to cause the change measurements in both the 5 days preceding and the
back to Status 1. Finally, the DNS functions in the 5 days following dfo = 0. There were 121 oestruses that
luteal phase and during pregnancy were modified to fulfilled these criteria. The average progesterone profile
better optimize the sampling frequency relative to time of confirmed oestruses is shown in Fig. 1.
of expected next oestrus (equations in Appendix).
Because sampling frequency in the present study was
not based on DNS, this change had no bearing on the Ratified oestruses
results presented here, except for those relating directly A high degree of consistency in the shape of the
to changing the sampling frequency. progesterone profiles of confirmed oestruses exists
(Fig. 1). This feature was used as a template to identify
segments of the progesterone profiles that match the
Model testing profile of the confirmed oestruses, hereafter called
The model has many features that warrant testing but in ratified oestruses, using the following profile matching
the present study we have chosen to focus on the procedure. By using the progesterone profiles from
detection of oestrus, end of postpartum anoestrus, and confirmed oestruses, the means (l) for each time point
onset of pregnancy, because these are the major outputs, (dfo = )5 to +10 in half-day steps) and the covariance
and those by which the usefulness of the model in matrix (P) were estimated. (To ensure that P is not ill-
commercial application will be primarily judged. For conditioned, the diagonal entries of P were multiplied by
this study, the model was run in its most basic form with 2 prior to further calculations). Then, for any given 15-
no additional inputs other than inseminations. day segment of a progesterone profile (x), the Maha-
lanobis distance D (Mahalanobis 1936) was calculated
as: D = ((x - l)TP)1(x - l))1 ⁄ 2 D will, under regularity
Oestrus detection conditions, have a v2 distribution with degrees of
Testing the ability of the model to identify oestrus freedom equal to the number of components in x. Large
implies that there are a series of known oestruses within values of D indicate that x is not consistent with l and
the data against which to compare the progesterone- P. Based on D, we calculated the corresponding tail-
based model detected oestruses. Both visual observa- probabilities p, and high values of p thus indicates that x
tions of oestrus and activity measurement-based oestrus is consistent with l and P. Log-transformed progester-
alarms were available on the test farm. However, it is one values were used because the error associated with
well known from other studies (Rossing and Spahr progesterone measurements is approximately propor-
1992), and was patently clear from inspection of the tional to the concentration of progesterone.
data, that these external oestrus detection methods were This basic procedure was carried out for a rolling 15-
not sufficiently reliable for the purpose of testing the day window through each individual progesterone
model. Accordingly, an alternative approach to defining profile (i.e. n ) 15 times for a profile of n daily
a reference set of known oestruses was made. Two types measurements). In order to convert the probabilities
of known oestrus were identified: confirmed and ratified into identifications of ratified oestrus, the following
oestruses. A confirmed oestrus was defined as an oestrus
at which insemination resulted in a confirmed preg-
28
nancy. A ratified oestrus is one in which the shape of the
progesterone profile matches that of the average pro- 24
gesterone profile of the confirmed oestruses.
Progesterone (ng/ml)
20
Confirmed oestruses 16
Inseminations that occurred within the time window 12
±10 days around the time-point 284 days before a
subsequent calving or 40 days before a positive preg- 8
nancy determination (rectal palpation 6 weeks post-
insemination) were assumed to be associated with a true, 4
confirmed oestrus (because pregnancy resulted). The 0
start of these confirmed oestruses was defined as the –10 –8 –6 –4 –2 0 2 4 6 8 10
time of the first smoothed milk progesterone measure- Days from onset of oestrus
ment <4 ng ⁄ ml (days from start of oestrus (dfo) = 0).
For those cows that subsequently calved, the calving Fig. 1. The average unsmoothed progesterone concentration (bold
date was used to define the time-window of conception line) of pregnancy-confirmed oestruses relative to days from onset of
oestrus. The 75% and 25% quartiles of the distribution of progester-
irrespective of any pregnancy determinations. Preg- one concentrations at each time-point are shown as dotted lines. Onset
nancy determination dates were only used to determine of oestrus was defined as the first decrease <4 ng ⁄ ml in smoothed
the time window of conception in those cows that had progesterone concentration
thresholds were applied. The probabilities had to be respectively. With each of these reduced data sets, the
greater than 0.95 and the weighted number of observa- model was run and the proportion of ratified oestruses
tions had to be greater than 3. The weighted number of that was detected by the model using the original data
observations was calculated as: set was calculated. In addition, the model was run using
X only those progesterone measurements that it would
(absðt tmidpoint Þ þ 1Þ1 ; have requested if it had been running in real-time
according to the model output DNS function. This is a
where t is time from the midpoint of the segment for any variable sampling frequency as described previously
given observation. For observations 7.5, 5, 2.5 and (and in Appendix). The total number of progesterone
0 days from the midpoint the individual weights were; samples used in this case was 11 957.
0.12, 0.17, 0.29 and 1, respectively. The threshold of
three for the sum of the individual weights corresponds Results and Discussion
to an observation frequency of two progesterone mea-
surements per 3 days for evenly spread observations, or Confirmed oestruses
a hole of )3 to +3 days from the midpoint in an Out of 121 confirmed oestruses, 104 were associated
otherwise complete set of observations. with a model-detected oestrus based on a decrease in
In order to avoid the same oestrus being repeatedly smoothed progesterone below 4 ng ⁄ ml. An additional
identified as the rolling 15-day window moves through 16 of the confirmed oestruses were model-detected on
the progesterone time series, the following filtering rules the basis of the supplementary oestrus detection rule
were applied. In addition to p being >0.95, p for the that smoothed progesterone decreased below 6 ng ⁄ ml
preceding observation had to be >0.90. Once an having previously been >15 ng ⁄ ml. Thus, the model as
observation was ratified, then subsequent observations implemented detected 99.2% of the confirmed oestruses,
meeting the above criteria were not accepted as ratified whereas the simpler model (4 ng ⁄ ml threshold) resulted
unless the probability had decreased to <0.5 in the in 85.6% of confirmed oestruses being detected.
interim. In addition, an observation was not ratified The majority of studies in the literature have used
when the progesterone concentration was >8 ng ⁄ ml 3 ng ⁄ ml as a threshold for detecting oestrus (Lamming
(Mahalanobis distance is scale invariant, i.e. it assesses and Bulman 1976). Thus, a discrepancy exists between
the shape of the segment and not the absolute level of this definition and the fact that insemination of cows
the segment). The total number of ratified oestruses was having these ‘high progesterone oestruses’ still resulted
679. The shape of the progesterone profile shown in in conception. A possible explanation for this discrep-
Fig. 1 is that which is associated with oestruses that ancy could be related to the smoothing of the proges-
occur after a prior period of luteal activity. Oestruses terone profiles. The very nature of smoothing is that it is
that occur in conjunction with the end of the post- resistant to extreme values. Indeed, this was one reason
partum anoestrus period do not match this profile and for using a threshold of 4 rather than 3 ng ⁄ ml in the
have thus been excluded from the test of the models’ model. So, if insufficient low progesterone values are
ability to detect oestrus. available, the smoothed profile may not decrease below
4 ng ⁄ ml even when individual observations are
<4 ng ⁄ ml. This would especially be the case when
Pregnancy determination observations are missing during the follicular phase.
The ability of the model to identify cows as pregnant This possibility was examined by comparing the number
following insemination was evaluated in two ways. By of observations for the low threshold (LThresh; thresh-
definition, the progesterone profile following confirmed old = 4 ng ⁄ ml) with high threshold (HThresh; thresh-
oestruses should always be identified by the model as the old = 6 ng ⁄ ml) types of model-detected confirmed
cow being in Status 2 (pregnant). The proportion of oestrus. Equal numbers of progesterone measurements
cases in which this was so was calculated, and any existed in the window )5 to +10 dfo (mean = 16). The
exceptions were examined. For those inseminations that same was true for a narrower ()2 to +3 dfo) window
did not result in a confirmed pregnancy (n = 375), the around the time of detected oestrus. Likewise, the
distribution of time intervals between insemination and weights used in the Mahalanobis procedure did not
model detected pregnancy failure (return from Status 2 differ, i.e. no difference was detected in the spread of
to 1) was examined. observations. Thus, no significant differences were
detected in the sampling frequency between LThresh
and HThresh model-detected confirmed oestruses. In
Sampling frequency contrast, the minimum unsmoothed progesterone con-
The effects of reducing the progesterone measurement centrations for the HThresh detected oestruses were
frequency on model performance were evaluated by greater (P < 0.001) than for the LThresh detected
generating new data sets from the original, full, proges- oestruses (2.4 vs 0.7 ng ⁄ ml, SE = 0.9). Across the
terone data in which a proportion of the original whole time window )5 to +10 dfo, the progesterone
progesterone measures were removed. This was done by concentration of HThresh detected oestruses was
imposing a minimum interval between consecutive 3.1 ng ⁄ ml greater than LThresh detected oestruses
progesterone measurements (within each cow-lactation (P < 0.001). The results of the present study indicate
time-series). For minimum intervals of 1, 2, 3, 4 and that cows whose smoothed progesterone profile does not
5 days, the resulting data sets contained: 42 459, 26 455, decrease below 4 ng ⁄ ml can conceive. Although all these
16 756, 8713 and 1535 progesterone measurements cows had a minimum unsmoothed progesterone level
<4 ng ⁄ ml, they had significantly greater progesterone >4 ng ⁄ ml (31 of 59 of false negatives). Of these, in 24
concentrations, both minimum and average, suggesting cases smoothed progesterone did not decrease below
that between cow variation occurs in the absolute 6 ng ⁄ ml. This high-progesterone oestrus increasingly
concentrations of the progesterone profile. This implies seems to be a real biological phenomenon (Fig. 2). Of
that simple threshold rules may not ultimately be the the remaining 28 false negatives, enlarging the window
optimal way of detecting oestrus when using progester- ()3 to + 5 dfo) around oestrus for aligning model
one profiles. detected and ratified oestruses by ±3 days of the oestrus
Confirmed oestruses permit calculation of model captured 21 cases. Accepting these 21 cases as a match
sensitivity (99.2%), the proportion of true oestruses reduced the number of false negatives to 38 resulting in a
detected by the model, but they have two limitations. model sensitivity of 93.3%. This sensitivity compared
Confirmed oestruses represent a small subset of the total very favourably with the sensitivities reported for other
number of oestruses because a maximum of one methods of oestrus detection (Firk et al. 2002), even
confirmed oestrus occurs per parity. Further, although when these methods are used in very favourable
conception provides proof of oestrus, the converse, lack circumstances such as high intensity visual oestrus
of conception, does not mean that the oestrus in detection (85.7% Van Eerdenburg 2006) or after oestrus
question was false. Thus, it is not possible to calculate synchronisation (91.3% for tail paint, 81.4% for
model specificity (false positives) using the confirmed pedometers Cavalieri et al. 2003). Only the combined
oestruses. Both of these limitations are overcome by use activity, milk yield, milk temperature and conductivity
of the ratified oestruses. model of de Mol et al. (1999) had a sensitivity (94%)
that matched that for the ratified oestruses. Although
this sensitivity was less than the 99% found using the
Ratified oestruses confirmed oestruses.
A total of 679 ratified oestruses were detected, of which In addition to enlarging the number of oestruses
62 were associated with the onset of oestrus cycles at the accepted as ‘true’, the profile matching procedure allows
end of postpartum anoestrus. These are discussed examination of those cases in which the model indicated
separately. There were also 22 cases occurring more oestrus in the absence of a ratified oestrus, and thus
than 5 days post-insemination, visual inspection allows calculation of model specificity. There were 171
revealed that 11 of 22 were caused by a prolonged model-detected oestruses (excluding first increases in
follicular phase often with missing observations allow- progesterone >4 ng ⁄ ml) that did not match a ratified
ing a very late ‘re-identification’ by the profile matching oestrus (i.e. false positives). Thus, the model had a
procedure. These were judged to have been falsely positive predictive value (proportion of model detected
identified as ratified oestruses and thus were also oestruses that are true oestruses) of 72.2%
excluded from the data used to test the ability of the [445 ⁄ (445 + 171)]. The total number of progesterone
model to detect oestrus. There were 606 ratified measures in the period of oestrus cycling (Status = 1
oestruses used, of which the model detected 445 using and profile matching weight <3) was 13 613 giving a
the 4 ng ⁄ ml rule and 83 using the high progesterone
oestrus rule, resulting in a sensitivity of 87.1%. As this
30
sensitivity of 87% is markedly lower than the sensitivity
calculated using confirmed oestruses, the 78 ratified
Progesterone (ng/ml)
basic specificity of 98.7%. However, this assumes that are obvious to the end user of the system. The predicted
for any given progesterone observation oestrus could be likelihood of a potential insemination succeeding also
indicated. This was not the case, because in the model, provides the end user with helpful information for
once oestrus was indicated, a waiting period of 5 days making a decision about whether or not to inseminate
was mandatory before a new oestrus can be indicated the cow. This has been explored further by Friggens and
(Friggens and Chagunda 2005). Allowing for this Løvendahl (2008).
waiting period, the total number of progesterone mea- In the model, the first increase in the smoothed
surements for which oestrus could be determined was progesterone profile above 4 ng ⁄ ml indicates the end of
2 723 resulting in a specificity of 93.7%. This specificity the postpartum anoestrus and coincidentally the first
is similar to that found by de Mol et al. (1999). oestrus. For those first oestruses that were ratified, the
Examination of the false positives showed that they average difference between the time of model detected
were largely associated with low progesterone concen- first oestrus and ratified oestrus was 1.5 days
trations fluctuating around the 4 ng ⁄ ml threshold (SD = 1.6). The distribution of the differences was left
(Fig. 3). Nearly 53% (90 ⁄ 171) of the false positives skewed by those cases in which a period of fluctuation
came at the end of ‘oestrus’ cycles in which the occurred around the 4-ng ⁄ ml threshold (Fig. 3). The
maximum progesterone concentration did not exceed median difference was 2.2 days. Thus, the model is in
10 ng ⁄ ml. Only in 56 out of the 171 cases had the good agreement with the ratified oestruses. However,
maximum progesterone concentration in the preceeding because the ‘Mahalanobis template’ was based on
cycle exceeded 15 ng ⁄ ml. Clearly, every time such a oestruses that occurred after a prior period of luteal
fluctuating profile decreased below 4 ng ⁄ ml, the model activity, it did not identify all first increases in proges-
indicated oestrus. This is the major drawback of a terone above the threshold. The distribution of these
simple threshold-based rule, which cannot be overcome model-detected durations of postpartum anoestrus was
by a simple smoothing of the progesterone profiles. The similar to that reported by Royal et al. (2002).
number of false positives could be reduced by simply
using a much greater threshold, thus increasing model
specificity. Unfortunately, a greater threshold has a Pregnancy determination
significant unwanted cost in terms of reducing model Of the 121 confirmed pregnancies, 108 remained in
sensitivity (de Mol et al. 2001). Instead, each time Status 2 (model-detected pregnant) from conception
oestrus is detected the model calculates a likelihood of a until at least positive pregnancy determination (by rectal
potential insemination succeeding. Because this calcula- palpation), the sensitivity of the model for detecting
tion includes information about the height and length of pregnancy was thus 89.3%. Of the 11 cases in which the
the preceding oestrus cycle, these false positive oestruses model falsely detected the cow as no longer pregnant, 2
are associated with a very low likelihood of insemination were due to gaps in the progesterone data, 5 failed
succeeding. Average values of the predicted likelihood because the model judged that the insemination was
of a potential insemination succeeding when the max- mistimed, and 6 were caused by decreases in progester-
imum progesterone concentration in the preceding cycle one concentration. These decreases were in some cases
did not exceed 5, 10 or 15 ng ⁄ ml were: 0.01, 0.17 and indistinguishable from those one would detect if the cow
0.45 respectively (on a scale 0–0.9). Thus, false positives was returning to oestrus following early embryo loss
(e.g. a sudden decrease to < 10 ng ⁄ ml) but these cows
remained pregnant and subsequently calved. Given
30 these types of cases, it is hard to envisage being able
to substantially improve pregnancy detection without
24 simultaneously increasing the number of false negatives.
Progesterone (ng/ml)
The present results indicate that after 23 days the model General considerations
reliably detected pregnancy failure. This is substantially This paper provides a test of one particular model for
earlier than the timing of reliable pregnancy determina- interpreting progesterone time-series data. In this test
tion by rectal palpation at 6 weeks post-insemination the model was run using limited inputs, for simplicity
(Peters and Ball 1995). the potential benefits of including additional non-pro-
gesterone information such as input from other external
Sampling frequency oestrus-detection methods (e.g. pedometers), body
energy status, and urea concentrations were not evalu-
The above results were generated from a data set in ated. Despite this, the model performed at least as well
which average time between collection of samples was as other oestrus-detection systems (Firk et al. 2002;
1.4 days (for the period 0–120 days from calving). Cavalieri et al. 2003; Van Eerdenburg 2006). In this
Because these results are expected to be influenced by context, it would have been relevant to compare this
sampling frequency, we examined the effect of reducing model with other models for interpreting progesterone
sampling frequency. As shown in Fig. 5, the relative profiles. At the time of writing, we could find no other
proportion of ratified oestruses detected by the model published progesterone models, thus comparison is
decreased with decreasing sampling frequency. These limited to specific aspects such as oestrus detection
results were achieved by simulating an evenly spread measured by indicators other than progesterone (virtu-
reduction in samples and thus sampling frequency. ally all tests use progesterone as the reference measure).
However, this contrasts with the way the model is set up Although the present model performed substantially
better in terms of sensitivity than the majority of
methods in the literature (Firk et al. 2002; Cavalieri et
1.0 al. 2003; Van Eerdenburg 2006), this is not in itself a
Proportion of oestruses detected
by a classical ELISA method. It may be expected that DNSLProp, DNSLRat, DNSLlag, SModRat, and SModT are con-
such a method is more precise than the types of stants with the following values: 0.25, )0.4, 5, 0.75 and 4, respectively.
measurement resulting from in situ biosensors. Con- The DNS function during pregnancy was modified in a similar way to
versely, milk samples used in the present study were concentrate sampling frequency around the time approximately
collected from cows milked robotically, resulting in 22 days post-insemination and reduce it thereafter:
variable inter milking intervals, and thus variable milk PregStepT = Cyclen + PregStepLag
fat content (Friggens and Rasmussen 2001). This
introduces a greater variability in the milk progester- MaxStepPreg ¼ ðTopPregStep BotPregStepÞ
one measurements (Waldmann et al. 1999) than would expðexpðPregStepRat ððRunTime AITimeÞ
be expected from milk samples collected more con- PregStepTÞÞÞ þ BotPregStep
ventionally. On balance, the results presented in this
study probably reflect what can be expected under DNS2def ¼ MaxStepPreg expðexpðDNSLRat
commercial conditions, but this remains to be quan-
ðDayFromNextOest DNSLlagÞÞÞ
tified.
TimeFromAIFun ¼ expð expðTfAIRat*((RunTime
Conclusions AITime) - TfAIT)))
Friggens NC, Chagunda MGG, Bjerring M, Ridder C, Rossing W, Spahr SL, 1992: Automation in machine milking.
Højsgaard S, Larsen T, 2007: Estimating degree of mastitis In: Bramley AJ, Dodd FH, Mein GA (eds.), Machine
from time-series measurements in milk: a test of a model Milking and Lactation. Queen City Printers Inc., Burling-
based on lactate dehydrogenase measurements. J Dairy Sci ton, pp. 311–342.
90, 5415–5427. Royal MD, Darwash AO, Flint APF, Webb R, Woolliams JA,
Friggens NC, Løvendahl P, 2008: The potential of on-farm Lamming GE, 2000: Declining fertility in dairy cattle:
fertility profiles: in-line progesterone and activity changes in traditional and endocrine parameters of fertility.
measurements. In: Friggens NC, Royal MD, Smith RF Anim Sci 70, 487–501.
(eds.), Fertility in Dairy Cows: Bridging the Gaps, BSAS Royal MD, Pryce JE, Woolliams JA, Flint APF, 2002: The
Occasional Publication. British Society of Animal Science, genetic relationship between commencement of luteal activ-
Edinburgh, UK (in press). ity and calving interval, body condition score, production,
Friggens NC, Rasmussen MD, 2001: Milk quality assessment and linear type traits in Holstein-Friesian dairy cattle. J
in automatic milking systems: accounting for the effects of Dairy Sci 85, 3071–3080.
variable intervals between milkings on milk composition. Simersky R, Swaczynova J, Morris DA, Franek M, Strnad M,
Livest Prod Sci 73, 45–54. 2007: Development of an ELISA-based kit for the on-farm
Lamming GE, Bulman DC, 1976: Use of milk progesterone determination of progesterone in milk. Vet Med 52, 19–28.
radioimmunoassay in diagnosis and treatment of subfertility Sreenan JM, Diskin MG, Morris DG, 2001: Embryo survival
in dairy cows. Br Vet J 132, 507–517. rate in cattle: a major limitation to the achievement of high
Lamming GE, Darwash AO, 1998: The use of milk proges- fertility. In: Diskin MG (ed), British Society of Animal
terone profiles to characterise components of subfertility in Science Occasional Publication, 26 (1). British Society of
milked dairy cows. Anim Reprod Sci 52, 175–190. Animal Science, Edinburgh, pp. 93–104.
Lattec, 2007: http://www.herdnavigator.com/pages/id33.asp. Van Eerdenburg FJCM, 2006: Estrus detection in dairy cattle:
Mahalanobis PC, 1936: On the generalised distance in statis- how to beat the bull. Vlaams Diergen Tijds 75, 61–69.
tics. Proc Natl Inst Sci India 12, 49–55. Waldmann A, 1993: Enzyme immunoassay (EIA) for milk
Opsomer G, Coryn M, Deluyker H, De Kruif A, 1998: An progesterone using a monoclonal antibody. Anim Reprod
analysis of ovarian dysfunction in high yielding dairy cows Sci 34, 19–30.
after calving based on progesterone profiles. Reprod Waldmann A, Ropstad E, Landsverk K, Sørensen K, Sølverød
Domest Anim 33, 193–204. L, Dahl E, 1999: Level and distribution of progesterone in
Peters AR, Ball PJH, 1995: Reproduction in Cattle, 2nd edn. bovine milk in relation to storage in the mammary gland.
Blackwell Science, Oxford, UK. Anim Reprod Sci 56, 79–91.
Prandi A, Messina M, Tondoio A, Motta M, 1999:
Correlation between reproductive efficiency, as determined Author’s address (for correspondence): NC Friggens, Faculty of
by new mathematical indexes, and the body condition score Agricultural Sciences, University of Aarhus, Research Centre Foulum,
in dairy cows. Theriogenology 52, 1251–1265. PO Box 50, 8830 Tjele, Denmark. E-mail: n.friggens@agrsci.dk
Romano JE, Thompson JA, Forrest DW, Westhusin ME,
Tomaszweski MA, Kraemer DC, 2006: Early pregnancy Conflict of interest: NC Friggens has received a research grant for this
diagnosis by transrectal ultrasonography in dairy cattle. work, part-funded by Lattec I ⁄ S; C Ridder is employed by Lattec I ⁄ S;
Theriogenology 66, 1034–1041. all remaining authors declare no conflict of interests.