Reprod Dom Anim 43 (Suppl. 2), 113–121 (2008); doi: 10.1111/j.1439-0531.2008.01150.
x
ISSN 0936-6768
Improved Detection of Reproductive Status in Dairy Cows Using Milk Progesterone
Measurements
NC Friggens1, M Bjerring1, C Ridder1,2, S Højsgaard1 and T Larsen1
1
Faculty of Agricultural Sciences, University of Aarhus, Research Centre Foulum, Tjele; 2Lattec I ⁄ S, Hillerød, Denmark
Contents
This study tested a model for predicting reproductive status
from in-line milk progesterone ‘measurements. The model is
that of Friggens and Chagunda [Theriogenology 64 (2005)
155]. Milk progesterone measurements (n = 55 036) repre-
senting 578 lactations from 380 cows were used to test the
model. Two types of known oestrus were identified: (1)
confirmed oestrus (at which insemination resulted in a
confirmed pregnancy, n = 121) and (2) ratified oestrus (where
the shape of the progesterone profile matched that of the
average progesterone profile of a confirmed oestrus, n = 679).
The model detected 99.2% of the confirmed oestruses. This
included a number of cases (n = 16) where the smoothed
progesterone did not decrease below 4 ng ⁄ ml. These cows had
significantly greater concentrations of progesterone, both
minimum and average, suggesting that between cow variation
exists in the absolute level of the progesterone profile. Using
ratified oestruses, model sensitivity was 93.3% and specificity
was 93.7% for detection of oestrus. Examination of false
positives showed that they were largely associated with low
concentrations of progesterone, fluctuating around the
4 ng ⁄ ml threshold. The distribution of time from insemination
until the model detected pregnancy failure had a median of
22 days post-insemination. In this test, the model was run
using limited inputs, the potential benefits of including
additional non-progesterone information were not evaluated.
Despite this, the model performed at least as well as other
oestrus detection systems.
Introduction
Progesterone measurements have been used for some
time now, and are accepted as a valid indicator, for
assessing the reproductive status of dairy cows (Bulman
and Lamming 1978; Royal et al. 2000). Implicit in this
process is a set of biological rules that are used to
convert the progesterone data into different reproduc-
tive statuses (e.g. postpartum anoestrus, luteal and
follicular phases of the oestrous cycle and pregnancy).
Traditionally, these rules were expert opinions usually
applied manually to progesterone data (e.g. Lamming
and Darwash 1998). As measuring technology has
advanced from the original radioimmunoassays,
through laboratory based ELISA methods (Waldmann
1993), to on-farm manual tests (Simersky et al. 2007),
biosensors (Delwiche et al. 2001), and ultimately
towards automated in-line systems for measuring milk
progesterone (Lattec 2007), these rules become both
explicit and fixed. This is because automated systems
necessarily incorporate a model to process and condense
the raw data, providing the end-user with the relevant
biological interpretation in a palatable form.
The fact that these biological rules are codified allows
them to be tested. Testing the ability of these rules to
 2008 The Authors. Journal compilation  2008 Blackwell Verlag
correctly categorize reproductive status becomes increas-
ingly important as automated in-line progesterone mon-
itoring systems are implemented on commercial farms.
Clearly, the value of a progesterone monitoring system
depends to a large extent upon how good the biological
model is. The objective of this study was to test the ability
of a model based on in-line milk progesterone measure-
ments to predict reproductive status. The model tested is
that of Friggens and Chagunda (2005).
Time-series models exist for detecting oestrus from
milk traits other than progesterone (de Mol et al. 1999;
Firk et al. 2003), and decision strategies for interpreting
progesterone data have been suggested (Lamming and
Darwash 1998; Opsomer et al. 1998; Prandi et al. 1999;
Delwiche et al. 2001). However, to our knowledge, the
only published full biological model to predict repro-
ductive status based on a time-series of milk progester-
one measurements is that of Friggens and Chagunda
(2005).
The normal procedure for testing a new predictor of a
given state is to compare it with some ‘gold standard’
reference measure. However, in the present case the
model is based on the measure, progesterone, which is
the accepted gold standard for assessing the reproduc-
tive status of dairy cows (Peters and Ball 1995; Cavalieri
et al. 2003). Any conventional test of progesterone
against another indicator (e.g. visually determined
oestrus or pedometers) rapidly reverses polarity, becom-
ing instead a test of the other indicator against
progesterone. Thus, testing of such a model requires a
non-conventional approach, which to a large extent
becomes an exploration of the shapes of the progester-
one profiles. Consequently, this study also characterizes
key aspects of progesterone profiles according to
reproductive status.
Materials and Methods
Test data collection
Milk progesterone measurements were made on pro-
portional whole milk samples collected from all milking
cows in one research herd (Danish Cattle Research
Centre) during the period 12 September 2002 to 30
September 2006. Cows were milked in a robotic milking
system with free traffic (mean no. visits ⁄ day = 2.4). One
progesterone measure was to be made daily during the
first 120 days from calving. For the remainder of
lactation, progesterone measurements were to be made
every second day. Actual average intervals between
progesterone samples before and after 120 days from
calving were: 1.4 and 2.4 days, respectively. Progesterone
114
was analysed using the Ridgeway ELISA-kit (Ridgeway
Science Ltd, Gloucestershire, UK). Milk samples were
pipetted, diluted and distributed using a Biomek 2000
(Laboratory Automation Workstation, Beckman Coul-
ter, Fullerton, CA, USA). Milk samples (25 ll, diluted
1 + 2 with water) were handled according to the
manufacturer’s instructions, however incubation R4
was increased to 1 h 30 min. Plates were read using a
spectrophotometer ⁄ fluorometer, Fluostar (BMG Lab-
technologies, Offenburg, Germany) (575 nm). Analyses
were performed in 96-well plates; two sets of seven
standards (0–30 ng ⁄ ml), locally made using milk from
an ovariectomized cow and ethanolic progesterone
solutions, and two sets of two control samples were
used for every analysis and plate. For the low and high
controls, the intra assay precision (CV%) was 14.9 and
1.4, respectively; the inter-assay precision (CV%) was
32.7 and 20.1, respectively; and the average inaccuracy
(bias) was +0.82 and )0.60 ng ⁄ ml, respectively.
The time series of milk progesterone measurements
was examined for gaps longer than 14 days with no
progesterone values. All data following such gaps were
excluded. Further, the entire cow-lactation was excluded
if there was insufficient data to identify the end of the
postpartum anoestrus period (i.e. the time series did not
start at calving and there were less than five measure-
ments before the cow was detected to have elevated
progesterone and resumed oestrous cycles). The result-
ing data represented information during 578 lactations
from 380 cows. The total number of milk progesterone
measurements was 55 036.
Timing of inseminations was based on visual
detection of oestrus backed up by activity meter
indications (DeLaval, Tumba, Sweden). Farm staff
had no access to the progesterone data. Visual
observations for oestrus signs were made daily at
06:00, 12:00 and 16:00, each period lasted 20 min. Cows
showing oestrus earlier than 35 days from calving were
not inseminated to that oestrus. All cows were
inseminated to oestrus regardless of other management
factors such as age, milk yield, and whether the cow was
designated for subsequent culling. Inseminations were
carried out the same day for oestruses detected at the
morning observation, and the next morning for
oestruses detected in the afternoon and evening.
The model being tested
The model being tested here has been described in detail
by Friggens and Chagunda (2005). Briefly, the model is
based on the cow always being in one of three
reproductive states (Status), these are: postpartum
anoestrus (Status 0), oestrus cycling (Status 1), and
potentially pregnant (Status 2). In the model these
statuses are mutually exclusive. The Status the cow was
found to be in by the previous run of the model is always
the Status at the start of the current run. By default, the
cow is assumed to be in Status 0 at calving. The model is
dynamic and deterministic, designed to run each time a
new trigger input occurs using both the current and
previous values. The main model inputs are a smoothed
(extended Kalman filter) progesterone concentration
and slope. Inseminations and pregnancy determinations
 2008 The Authors. Journal compilation  2008 Blackwell Verlag
NC Friggens, M Bjerring, C Ridder, S Højsgaard and T Larsen
are also included as inputs. To make use of other known
effectors of reproductive performance that are not
necessarily reflected in progesterone concentrations,
the model is designed to incorporate a number of
additional inputs. However, the model is designed to
function in the absence of these additional inputs
(although with some loss of accuracy in some outputs).
The model outputs are all reproductive Status specific
with the exception of days to next sample (DNS) which
is calculated in each model run regardless of reproduc-
tive Status. Days to next sample is an important output
with respect to the progesterone measurements being
automated. It is designed to feedback to the sampling
system so that the frequency of milk sampling (i.e.
progesterone measurement) can be varied according to
the predicted likelihood of a future reproductive event
such as onset of oestrus cycles. The other model outputs
are: risk of prolonged postpartum anoestrus, risk and
type of ovarian cyst (i.e. prolonged luteal or follicular
phase), onset of oestrus, likelihood of a potential
insemination succeeding and likelihood of being preg-
nant (following oestrus).
The shift from Status 0 (postpartum anoestrus) to
Status 1 (oestrus cycling) is caused by two consecutive
smoothed progesterone measurements above the thresh-
old value (LThresh) which is 4 ng ⁄ ml. (If EOD inputs
are available these can also provoke this shift). As
described by Friggens and Chagunda (2005), once a cow
is in Status 1 the shift to Status 2 is caused by the
smoothed progesterone value decreasing below
LThresh. This shift generates a model-detected indica-
tion of oestrus. If no insemination is recorded in the
following 5 days, the cow automatically reverts to Status
1. If there is a timely insemination and subsequent
progesterone concentrations increase, then the cow
remains in status 2 (potentially pregnant) as long as
the model predicted likelihood of pregnancy remains
above a pregnancy threshold. Typically, if progesterone
concentrations decrease, then the model shifts back
from Status 2 to Status 1.
The model tested in the present study contained the
following major modification relative to the model
described by Friggens and Chagunda (2005). Visual
inspection of the earliest progesterone profiles collected
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-
duced (6 ng ⁄ ml). This allowed identification of these
‘high-progesterone oestruses’ conditional on the maxi-
mum progesterone concentration in the preceding cycle
being greater than 15 ng ⁄ ml, and the present decrease
being >10 days since a preceding oestrus.
The model tested in the present study also contained
the following minor modifications relative to the model
described by Friggens and Chagunda (2005). The slope
of the progesterone profile is used, for example in the
function calculating the predicted likelihood of AI
success. The slope can have extreme values particularly
with short intervals between measurements, thus an
adjusted slope (a sigmoid transformation) was used
instead. Over the normal range this makes virtually no
difference but it caps the extremes. In Status 2, when the
calculated likelihood of pregnancy (LikePreg) decreases
Reproductive Status Assessed by Milk Progesterone
below a pregnancy threshold, the cow is deemed no
longer pregnant and reverts to Status 1. In the current
implementation of the model, a modification was made
such that once the cow was more than 30 days after
insemination, 2 consecutive low LikePreg (i.e. low
progesterone) values are required to cause the change
back to Status 1. Finally, the DNS functions in the
luteal phase and during pregnancy were modified to
better optimize the sampling frequency relative to time
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
results presented here, except for those relating directly
to changing the sampling frequency.
Model testing
The model has many features that warrant testing but in
the present study we have chosen to focus on the
detection of oestrus, end of postpartum anoestrus, and
onset of pregnancy, because these are the major outputs,
and those by which the usefulness of the model in
commercial application will be primarily judged. For
this study, the model was run in its most basic form with
no additional inputs other than inseminations.
Oestrus detection
Testing the ability of the model to identify oestrus
implies that there are a series of known oestruses within
the data against which to compare the progesterone-
based model detected oestruses. Both visual observa-
tions of oestrus and activity measurement-based oestrus
alarms were available on the test farm. However, it is
well known from other studies (Rossing and Spahr
1992), and was patently clear from inspection of the
data, that these external oestrus detection methods were
not sufficiently reliable for the purpose of testing the
model. Accordingly, an alternative approach to defining
a reference set of known oestruses was made. Two types
of known oestrus were identified: confirmed and ratified
oestruses. A confirmed oestrus was defined as an oestrus
at which insemination resulted in a confirmed preg-
nancy. A ratified oestrus is one in which the shape of the
progesterone profile matches that of the average pro-
gesterone profile of the confirmed oestruses.
Confirmed oestruses
Inseminations that occurred within the time window
±10 days around the time-point 284 days before a
subsequent calving or 40 days before a positive preg-
nancy determination (rectal palpation 6 weeks post-
insemination) were assumed to be associated with a true,
confirmed oestrus (because pregnancy resulted). The
start of these confirmed oestruses was defined as the
time of the first smoothed milk progesterone measure-
ment <4 ng ⁄ ml (days from start of oestrus (dfo) = 0).
For those cows that subsequently calved, the calving
date was used to define the time-window of conception
irrespective of any pregnancy determinations. Preg-
nancy determination dates were only used to determine
the time window of conception in those cows that had
 2008 The Authors. Journal compilation  2008 Blackwell Verlag
115
not calved by the end of the experiment period. When
more than one insemination occurred within a time
window, then the last one was chosen. For the purpose
of testing the progesterone model, it was required that
confirmed oestruses had more than 2 progesterone
measurements in both the 5 days preceding and the
5 days following dfo = 0. There were 121 oestruses that
fulfilled these criteria. The average progesterone profile
of confirmed oestruses is shown in Fig. 1.
Ratified oestruses
A high degree of consistency in the shape of the
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
profile of the confirmed oestruses, hereafter called
ratified oestruses, using the following profile matching
procedure. By using the progesterone profiles from
confirmed oestruses, the means (l) for each time point
(dfo = )5 to +10 in half-day steps) and the covariance
matrix (P) were estimated. (To ensure that P is not ill-
conditioned, the diagonal entries of P were multiplied by
2 prior to further calculations). Then, for any given 15-
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
conditions, have a v2 distribution with degrees of
freedom equal to the number of components in x. Large
values of D indicate that x is not consistent with l and
P. Based on D, we calculated the corresponding tail-
probabilities p, and high values of p thus indicates that x
is consistent with l and P. Log-transformed progester-
one values were used because the error associated with
progesterone measurements is approximately propor-
tional to the concentration of progesterone.
This basic procedure was carried out for a rolling 15-
day window through each individual progesterone
profile (i.e. n ) 15 times for a profile of n daily
measurements). In order to convert the probabilities
into identifications of ratified oestrus, the following
28
24
Progesterone (ng/ml)
20
16
12
8
4
0
–10 –8 –6 –4 –2 0 2 4 6 8 10
Days from onset of oestrus
Fig. 1. The average unsmoothed progesterone concentration (bold
line) of pregnancy-confirmed oestruses relative to days from onset of
oestrus. The 75% and 25% quartiles of the distribution of progester-
one concentrations at each time-point are shown as dotted lines. Onset
of oestrus was defined as the first decrease <4 ng ⁄ ml in smoothed
progesterone concentration
116
thresholds were applied. The probabilities had to be
greater than 0.95 and the weighted number of observa-
tions had to be greater than 3. The weighted number of
observations was calculated as:
X only those progesterone measurements that it would
(absðt  tmidpoint Þ þ 1Þ1 ;
where t is time from the midpoint of the segment for any
given observation. For observations 7.5, 5, 2.5 and
0 days from the midpoint the individual weights were;
0.12, 0.17, 0.29 and 1, respectively. The threshold of
three for the sum of the individual weights corresponds
to an observation frequency of two progesterone mea-
surements per 3 days for evenly spread observations, or
a hole of )3 to +3 days from the midpoint in an
otherwise complete set of observations.
In order to avoid the same oestrus being repeatedly
identified as the rolling 15-day window moves through
the progesterone time series, the following filtering rules
were applied. In addition to p being >0.95, p for the
preceding observation had to be >0.90. Once an
observation was ratified, then subsequent observations
meeting the above criteria were not accepted as ratified
unless the probability had decreased to <0.5 in the
interim. In addition, an observation was not ratified
when the progesterone concentration was >8 ng ⁄ ml
(Mahalanobis distance is scale invariant, i.e. it assesses
the shape of the segment and not the absolute level of
the segment). The total number of ratified oestruses was
679. The shape of the progesterone profile shown in
Fig. 1 is that which is associated with oestruses that
occur after a prior period of luteal activity. Oestruses
that occur in conjunction with the end of the post-
partum anoestrus period do not match this profile and
have thus been excluded from the test of the models’
ability to detect oestrus.
Pregnancy determination
The ability of the model to identify cows as pregnant
following insemination was evaluated in two ways. By
definition, the progesterone profile following confirmed
oestruses should always be identified by the model as the
cow being in Status 2 (pregnant). The proportion of
cases in which this was so was calculated, and any
exceptions were examined. For those inseminations that
did not result in a confirmed pregnancy (n = 375), the
distribution of time intervals between insemination and
model detected pregnancy failure (return from Status 2
to 1) was examined.
Sampling frequency
The effects of reducing the progesterone measurement
frequency on model performance were evaluated by
generating new data sets from the original, full, proges-
terone data in which a proportion of the original
progesterone measures were removed. This was done by
imposing a minimum interval between consecutive
progesterone measurements (within each cow-lactation
time-series). For minimum intervals of 1, 2, 3, 4 and
5 days, the resulting data sets contained: 42 459, 26 455,
16 756, 8713 and 1535 progesterone measurements
 2008 The Authors. Journal compilation  2008 Blackwell Verlag
NC Friggens, M Bjerring, C Ridder, S Højsgaard and T Larsen
respectively. With each of these reduced data sets, the
model was run and the proportion of ratified oestruses
that was detected by the model using the original data
set was calculated. In addition, the model was run using
have requested if it had been running in real-time
according to the model output DNS function. This is a
variable sampling frequency as described previously
(and in Appendix). The total number of progesterone
samples used in this case was 11 957.
Results and Discussion
Confirmed oestruses
Out of 121 confirmed oestruses, 104 were associated
with a model-detected oestrus based on a decrease in
smoothed progesterone below 4 ng ⁄ ml. An additional
16 of the confirmed oestruses were model-detected on
the basis of the supplementary oestrus detection rule
that smoothed progesterone decreased below 6 ng ⁄ ml
having previously been >15 ng ⁄ ml. Thus, the model as
implemented detected 99.2% of the confirmed oestruses,
whereas the simpler model (4 ng ⁄ ml threshold) resulted
in 85.6% of confirmed oestruses being detected.
The majority of studies in the literature have used
3 ng ⁄ ml as a threshold for detecting oestrus (Lamming
and Bulman 1976). Thus, a discrepancy exists between
this definition and the fact that insemination of cows
having these ‘high progesterone oestruses’ still resulted
in conception. A possible explanation for this discrep-
ancy could be related to the smoothing of the proges-
terone profiles. The very nature of smoothing is that it is
resistant to extreme values. Indeed, this was one reason
for using a threshold of 4 rather than 3 ng ⁄ ml in the
model. So, if insufficient low progesterone values are
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
observations are missing during the follicular phase.
This possibility was examined by comparing the number
of observations for the low threshold (LThresh; thresh-
old = 4 ng ⁄ ml) with high threshold (HThresh; thresh-
old = 6 ng ⁄ ml) types of model-detected confirmed
oestrus. Equal numbers of progesterone measurements
existed in the window )5 to +10 dfo (mean = 16). The
same was true for a narrower ()2 to +3 dfo) window
around the time of detected oestrus. Likewise, the
weights used in the Mahalanobis procedure did not
differ, i.e. no difference was detected in the spread of
observations. Thus, no significant differences were
detected in the sampling frequency between LThresh
and HThresh model-detected confirmed oestruses. In
contrast, the minimum unsmoothed progesterone con-
centrations for the HThresh detected oestruses were
greater (P < 0.001) than for the LThresh detected
oestruses (2.4 vs 0.7 ng ⁄ ml, SE = 0.9). Across the
whole time window )5 to +10 dfo, the progesterone
concentration of HThresh detected oestruses was
3.1 ng ⁄ ml greater than LThresh detected oestruses
(P < 0.001). The results of the present study indicate
that cows whose smoothed progesterone profile does not
decrease below 4 ng ⁄ ml can conceive. Although all these
cows had a minimum unsmoothed progesterone level
Reproductive Status Assessed by Milk Progesterone
<4 ng ⁄ ml, they had significantly greater progesterone
concentrations, both minimum and average, suggesting
that between cow variation occurs in the absolute
concentrations of the progesterone profile. This implies
that simple threshold rules may not ultimately be the
optimal way of detecting oestrus when using progester-
one profiles.
Confirmed oestruses permit calculation of model
sensitivity (99.2%), the proportion of true oestruses
detected by the model, but they have two limitations.
Confirmed oestruses represent a small subset of the total
number of oestruses because a maximum of one
confirmed oestrus occurs per parity. Further, although
conception provides proof of oestrus, the converse, lack
of conception, does not mean that the oestrus in
question was false. Thus, it is not possible to calculate
model specificity (false positives) using the confirmed
oestruses. Both of these limitations are overcome by use
of the ratified oestruses.
Ratified oestruses
A total of 679 ratified oestruses were detected, of which
62 were associated with the onset of oestrus cycles at the
end of postpartum anoestrus. These are discussed
separately. There were also 22 cases occurring more
than 5 days post-insemination, visual inspection
revealed that 11 of 22 were caused by a prolonged
follicular phase often with missing observations allow-
ing a very late ‘re-identification’ by the profile matching
procedure. These were judged to have been falsely
identified as ratified oestruses and thus were also
excluded from the data used to test the ability of the
model to detect oestrus. There were 606 ratified
oestruses used, of which the model detected 445 using
the 4 ng ⁄ ml rule and 83 using the high progesterone
oestrus rule, resulting in a sensitivity of 87.1%. As this
sensitivity of 87% is markedly lower than the sensitivity
calculated using confirmed oestruses, the 78 ratified
oestruses that were not detected by the model were
examined further. Of the 78 ratified oestruses that
were not detected by the model (i.e. false negatives), 19
were considered to be a detection failure of the profile
matching procedure (primarily associated with pro-
longed follicular phase) rather then a detection failure
of the model. Excluding these 19 cases reduced the
number of false negatives to 59 and consequently
increased model sensitivity to 89.9%.
Making this visual adjustment of the procedure for
identifying ratified oestruses may be viewed as inappro-
priate. However, it can also be viewed as reflecting the
fact that inherent in any automated procedure to detect
features of a profile of real, and thus noisy data, there
will always be cases where the procedure fails. Because
the most reliable measure of the model’s sensitivity is the
one calculated from the confirmed oestruses, and given
that there is no a priori reason for expecting the true
sensitivity to be different for ratified oestruses, this latter
justification seems reasonable (for completeness, all the
different sensitivities have been presented). As with the
confirmed oestruses, the main reason for false negatives
(i.e. the model not detecting true ratified oestruses) is
that these oestruses all had minima substantially
 2008 The Authors. Journal compilation  2008 Blackwell Verlag
117
>4 ng ⁄ ml (31 of 59 of false negatives). Of these, in 24
cases smoothed progesterone did not decrease below
6 ng ⁄ ml. This high-progesterone oestrus increasingly
seems to be a real biological phenomenon (Fig. 2). Of
the remaining 28 false negatives, enlarging the window
()3 to + 5 dfo) around oestrus for aligning model
detected and ratified oestruses by ±3 days of the oestrus
captured 21 cases. Accepting these 21 cases as a match
reduced the number of false negatives to 38 resulting in a
model sensitivity of 93.3%. This sensitivity compared
very favourably with the sensitivities reported for other
methods of oestrus detection (Firk et al. 2002), even
when these methods are used in very favourable
circumstances such as high intensity visual oestrus
detection (85.7% Van Eerdenburg 2006) or after oestrus
synchronisation (91.3% for tail paint, 81.4% for
pedometers Cavalieri et al. 2003). Only the combined
activity, milk yield, milk temperature and conductivity
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
confirmed oestruses.
In addition to enlarging the number of oestruses
accepted as ‘true’, the profile matching procedure allows
examination of those cases in which the model indicated
oestrus in the absence of a ratified oestrus, and thus
allows calculation of model specificity. There were 171
model-detected oestruses (excluding first increases in
progesterone >4 ng ⁄ ml) that did not match a ratified
oestrus (i.e. false positives). Thus, the model had a
positive predictive value (proportion of model detected
oestruses that are true oestruses) of 72.2%
[445 ⁄ (445 + 171)]. The total number of progesterone
measures in the period of oestrus cycling (Status = 1
and profile matching weight <3) was 13 613 giving a
30
Progesterone (ng/ml)
24
18
12
6
0
0 30 60 90 120 150 180
Days from calving
Fig. 2. An example of a progesterone profile showing a very high
progesterone oestrus. Because the smoothed progesterone profile (solid
line) did not decrease below 6 ng/ml, the model failed to detect this
oestrus (at day 92). Subsequent constantly elevated progesterone and a
positive pregnancy determination by rectal palpation (at day 128)
indicated that this cow conceived despite the very high concentration
of progesterone at oestrus. Raw, unsmoothed progesterone values are
shown by the open circles, the smoothed progesterone profile is
shown by the solid line. Pluses indicate model-determined Status 0
(postpartum an oestrus), solid triangles indicate Status 1 (oestrus
cycling), and solid squares indicate Status 2 (potentially pregnant).
The downward pointing open triangles indicate inseminations, the
upwards pointing open triangles indicate pregnancy determinations
(1 ‘‘ng/ml’’ = not pregnant, 9 ‘‘ng/ml’’ = pregnant)
118
basic specificity of 98.7%. However, this assumes that
for any given progesterone observation oestrus could be
indicated. This was not the case, because in the model,
once oestrus was indicated, a waiting period of 5 days
was mandatory before a new oestrus can be indicated
(Friggens and Chagunda 2005). Allowing for this
waiting period, the total number of progesterone mea-
surements for which oestrus could be determined was
2 723 resulting in a specificity of 93.7%. This specificity
is similar to that found by de Mol et al. (1999).
Examination of the false positives showed that they
were largely associated with low progesterone concen-
trations fluctuating around the 4 ng ⁄ ml threshold
(Fig. 3). Nearly 53% (90 ⁄ 171) of the false positives
came at the end of ‘oestrus’ cycles in which the
maximum progesterone concentration did not exceed
10 ng ⁄ ml. Only in 56 out of the 171 cases had the
maximum progesterone concentration in the preceeding
cycle exceeded 15 ng ⁄ ml. Clearly, every time such a
fluctuating profile decreased below 4 ng ⁄ ml, the model
indicated oestrus. This is the major drawback of a
simple threshold-based rule, which cannot be overcome
by a simple smoothing of the progesterone profiles. The
number of false positives could be reduced by simply
using a much greater threshold, thus increasing model
specificity. Unfortunately, a greater threshold has a
significant unwanted cost in terms of reducing model
sensitivity (de Mol et al. 2001). Instead, each time
oestrus is detected the model calculates a likelihood of a
potential insemination succeeding. Because this calcula-
tion includes information about the height and length of
the preceding oestrus cycle, these false positive oestruses
are associated with a very low likelihood of insemination
succeeding. Average values of the predicted likelihood
of a potential insemination succeeding when the max-
imum progesterone concentration in the preceding cycle
did not exceed 5, 10 or 15 ng ⁄ ml were: 0.01, 0.17 and
0.45 respectively (on a scale 0–0.9). Thus, false positives
30
24
Progesterone (ng/ml)
18
12
6 insemination to model-detected pregnancy failure is
0 assessing the cow to not be pregnant: (1) inappropriate
0 30 60 90 120 150 180
Days from calving
Fig. 3. An example of a progesterone profile showing a fluctuating low
progesterone after calving. Because the smoothed progesterone profile
(solid line) repeatedly decreases below 4 ng/ml during this period, the
model repeatedly detected false positive oestruses. Raw, unsmoothed
progesterone values are shown by the open circles. Pluses indicate
model-determined Status 0 (postpartum an oestrus), solid triangles
indicate Status 1 (oestrus cycling), and solid squares indicate Status 2
(potentially pregnant). The downward pointing open triangles indicate
inseminations, the upwards pointing open triangles indicate pregnancy
determinations (1 ‘‘ng/ml’’ = not pregnant, 9 ‘‘ng/ml’’ = pregnant)
 2008 The Authors. Journal compilation  2008 Blackwell Verlag
NC Friggens, M Bjerring, C Ridder, S Højsgaard and T Larsen
are obvious to the end user of the system. The predicted
likelihood of a potential insemination succeeding also
provides the end user with helpful information for
making a decision about whether or not to inseminate
the cow. This has been explored further by Friggens and
Løvendahl (2008).
In the model, the first increase in the smoothed
progesterone profile above 4 ng ⁄ ml indicates the end of
the postpartum anoestrus and coincidentally the first
oestrus. For those first oestruses that were ratified, the
average difference between the time of model detected
first oestrus and ratified oestrus was 1.5 days
(SD = 1.6). The distribution of the differences was left
skewed by those cases in which a period of fluctuation
occurred around the 4-ng ⁄ ml threshold (Fig. 3). The
median difference was 2.2 days. Thus, the model is in
good agreement with the ratified oestruses. However,
because the ‘Mahalanobis template’ was based on
oestruses that occurred after a prior period of luteal
activity, it did not identify all first increases in proges-
terone above the threshold. The distribution of these
model-detected durations of postpartum anoestrus was
similar to that reported by Royal et al. (2002).
Pregnancy determination
Of the 121 confirmed pregnancies, 108 remained in
Status 2 (model-detected pregnant) from conception
until at least positive pregnancy determination (by rectal
palpation), the sensitivity of the model for detecting
pregnancy was thus 89.3%. Of the 11 cases in which the
model falsely detected the cow as no longer pregnant, 2
were due to gaps in the progesterone data, 5 failed
because the model judged that the insemination was
mistimed, and 6 were caused by decreases in progester-
one concentration. These decreases were in some cases
indistinguishable from those one would detect if the cow
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
these types of cases, it is hard to envisage being able
to substantially improve pregnancy detection without
simultaneously increasing the number of false negatives.
However, of greater important to the end user is
detecting true pregnancy failure.
The distribution of model detected pregnancy lengths,
for those pregnancies that did not result in a positive
pregnancy determination or a calf, provides useful
information (Fig. 4). The distribution of time from
bimodal reflecting the two reasons for the model
timing of insemination and (2) progesterone concentra-
tions too low (first and second peaks, respectively). The
median of the distribution was 22 days, reflecting the
second peak. By 27 days 60% of pregnancy failures had
occurred, and by 34 days 70% had occurred. These
figures reflect a typical time pattern of early embryonic
loss (Sreenan et al. 2001). Reliable detection of preg-
nancy using progesterone is not possible much earlier
than 18–23 days post-insemination because the proges-
terone profile during the luteal phase is not affected by
presence of an embryo (Bulman and Lamming 1978).
Reproductive Status Assessed by Milk Progesterone 119

0.16 to request samples. In a situation in which the model is

part of an automated in-line progesterone measuring
0.14
system, it controls sampling frequency through the DNS
Proportion of cases

0.12 function. The DNS function implements a variable

0.10
sampling frequency that is low immediately after
oestrus, and then increases as a function of both
0.08 duration of luteal phase and any evidence of a decrease
0.06 in progesterone (Friggens and Chagunda 2005; see also
Appendix). Using this variable function to reduce the
0.04 sampling frequency in the test data resulted in an
0.02 average sampling frequency of 3.3 days, but the pro-
portional reduction in oestrus detection was only 0.73.
0.00 This is notably smaller than the 0.45 reduction for the
10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170
Days from calving equivalent equally distributed sampling frequency
(Fig. 5). Even in this conservative test of sampling
Fig. 4. Distribution of pregnancy lengths for cows in which pregnancy frequency effects there is a clear benefit of having a
failed (model detected reproductive status reverted from pregnancy to variable sampling frequency.
oestrus cycling)

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

definitive assessment of the models quality. As demon-

0.8 strated by numerous authors, sensitivity and specificity
can be altered by varying the thresholds used, and
generally have an inverse relation to each other (de Mol
0.6 et al. 2001; Faustini et al. 2007). In the context of time-
series measurements such as progesterone profiles,
0.4 sensitivity and specificity measures are particularly
unsatisfactory because they are calculated for fixed
time-points or windows (Friggens et al. 2007). For
0.2
example, Faustini et al. (2007) reported a sensitivity for
progesterone determination of pregnancy of 98.2%,
0.0 substantially greater than the 89.3% reported in the
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 present study, whereas Romano et al. (2006) reported a
Interval between progesterone measurements sensitivity of 74.5%. The difference was in large part
Fig. 5. The effect of reducing progesterone sampling frequency
because of the time (days post-insemination) at which
(increasing the average interval between progesterone measurements the sensitivity was calculated. Indeed, Romano et al.
[units = days]) on the proportion of ratified oestruses detected by the (2006) reported increasing sensitivity with increasing
model. The proportion is expressed relative to the case with the full days post-insemination. Clearly, factors such as the
progesterone data set. The solid line indicates the effect of reduced chosen time-point and sampling frequency make direct
sampling frequency using fixed sampling schedules. The solid circle
indicates the proportion of oestruses detected when using the model comparison of methods difficult.
output days to next sample function that results in a variable sampling It should also be noted that in the present study,
frequency progesterone measurements were made in a laboratory

 2008 The Authors. Journal compilation  2008 Blackwell Verlag

120
by a classical ELISA method. It may be expected that
such a method is more precise than the types of
measurement resulting from in situ biosensors. Con-
versely, milk samples used in the present study were
collected from cows milked robotically, resulting in
variable inter milking intervals, and thus variable milk
fat content (Friggens and Rasmussen 2001). This
introduces a greater variability in the milk progester-
one measurements (Waldmann et al. 1999) than would
be expected from milk samples collected more con-
ventionally. On balance, the results presented in this
study probably reflect what can be expected under
commercial conditions, but this remains to be quan-
tified.
Conclusions
The type of model needed to condense and interpret
progesterone profile data in real-time on-farm has been
tested and found to perform substantially better in terms
of sensitivity of oestrus detection than other existing
detection systems. The model also was shown to provide
valuable information about other aspects of reproduc-
tive status such as pregnancy determination and com-
mencement of luteal activity.
Acknowledgements
We gratefully acknowledge the valuable contribution of the farm staff
at KFC, Jens Clausen, Carsten Berthelsen, Peter Løvendahl, Jes
Nielsen, and Connie Middelhede for their efforts in securing this
massive number of samples. This study, which was part of the Biosens
project, was funded by the Danish Ministry of Food, Agriculture and
Fisheries and the Danish Cattle Association.
Appendix
Days to next sampling functions for the luteal phase and during
pregnancy
The original days to next sampling (DNS) function in the luteal
phase (Friggens and Chagunda 2005) decreased sampling frequency
as duration of oestrus cycle (cyclen) increased towards 21 but then
stayed low in prolonged luteal phases. The function was changed in
two ways. The default luteal DNS function (DNS1LdefR) was
changed from a declining sigmoid to a declining sigmoid plus a
subsequent climbing sigmoid (i.e. it rises again after expected
cyclen). The slope modifier was adjusted to use the absolute
difference between the maximum progesterone concentration since
last oestrus (CycMax) and current progesterone concentration rather
than the slope:
DayFromNextOest = abs((Date - DayOest) - Cyclen)
MaxStep ¼ Cyclen  DNSLProp
DNS1Ldef ¼ MaxStep  expðexpðDNSLRat
 ðDayFromNextOest  DNSLlagÞÞÞ
SlopeMod ¼ expð expðSModRat*((CycMax - Level)
 (CycMax/SModT))))
DNS ¼ DNS1Ldef  SlopeMod
where date is the current day, DayOest is the day of the preceding
oestrus, and level is the smoothed progesterone concentration (ng ⁄ ml).
 2008 The Authors. Journal compilation  2008 Blackwell Verlag
NC Friggens, M Bjerring, C Ridder, S Højsgaard and T Larsen
DNSLProp, DNSLRat, DNSLlag, SModRat, and SModT are con-
stants with the following values: 0.25, )0.4, 5, 0.75 and 4, respectively.
The DNS function during pregnancy was modified in a similar way to
concentrate sampling frequency around the time approximately
22 days post-insemination and reduce it thereafter:
PregStepT = Cyclen + PregStepLag
MaxStepPreg ¼ ðTopPregStep  BotPregStepÞ
 expðexpðPregStepRat  ððRunTime  AITimeÞ
 PregStepTÞÞÞ þ BotPregStep
DNS2def ¼ MaxStepPreg  expðexpðDNSLRat
 ðDayFromNextOest  DNSLlagÞÞÞ
TimeFromAIFun ¼ expð expðTfAIRat*((RunTime
 AITime) - TfAIT)))
offsetLevTime ¼ expðexpðPLevRat  ðLevelðPLevT
þ DNS2LevToffsetÞ  TimeFromAIFunRÞÞÞ
DNS2LevTime ¼ ð1  ð1  TimeFromAIFunÞÞ
 offsetLevTime þ ð1  TimeFromAIFunÞ
DNSR ¼ DNS2def  DNS2LevTime
where RunTime is the current time and AITime is the time of
insemination. PregStepLag, TopPregStep, BotPregStep, PregStepRat,
TfAIRat, TfAIT, PLevRat, PLevT and DNS2LevToffset are constants
with the following values: 6, 10, 5, )0.4, )0.3, 12, )0.5, 12 and 6,
respectively.
References
Bulman DC, Lamming GE, 1978: Milk progesterone levels
in relation to conception, repeat breeding and factors
influencing acyclicity in dairy cows. J Reprod Fertil 54,
447–458.
Cavalieri J, Eagles VE, Ryan M, Macmillan KL, 2003:
Comparison of four methods for detection of oestrus in
dairy cows with resynchronised oestrous cycles. Aust Vet J
81, 422–425.
de Mol RM, Keen A, Kroeze GH, Achten JMFH, 1999:
Description of a detection model for oestrus and diseases in
dairy cattle based on time series analysis combined with a
Kalman filter. Comput Electron Agr 22, 171–185.
de Mol RM, Ouweltjes W, Kroeze GH, Hendriks MMWB,
2001: Detection of estrus and mastitis: field performance of
a model. Appl Eng Agr 17, 399–407.
Delwiche MJ, Tang X, BonDurant RH, Munro CJ, 2001:
Estrus detection with a progesterone biosensor. Trans
ASAE 44, 2003–2008.
Faustini M, Battocchio M, Vigo D, Prandi A, Veronesi MC,
Comin A, Cairoli F, 2007: Pregnancy diagnosis in dairy
cows by whey progesterone analysis: an ROC approach.
Theriogenology 67, 1386–1392.
Firk R, Stamer E, Junge W, Krieter J, 2002: Automation of
oestrus detection in dairy cows: a review. Livest Prod Sci 75,
219–232.
Firk R, Stamer E, Junge W, Krieter J, 2003: Improving oestrus
detection by combination of activity measurements with
information about previous oestrus cases. Livest Prod Sci
82, 97–103.
Friggens NC, Chagunda MGG, 2005: Prediction of the
reproductive status of cattle on the basis of milk progester-
one measures: model description. Theriogenology 64,
155–190.
Reproductive Status Assessed by Milk Progesterone
Friggens NC, Chagunda MGG, Bjerring M, Ridder C,
Højsgaard S, Larsen T, 2007: Estimating degree of mastitis
from time-series measurements in milk: a test of a model
based on lactate dehydrogenase measurements. J Dairy Sci
90, 5415–5427.
Friggens NC, Løvendahl P, 2008: The potential of on-farm
fertility profiles: in-line progesterone and activity
measurements. In: Friggens NC, Royal MD, Smith RF
(eds.), Fertility in Dairy Cows: Bridging the Gaps, BSAS
Occasional Publication. British Society of Animal Science,
Edinburgh, UK (in press).
Friggens NC, Rasmussen MD, 2001: Milk quality assessment
in automatic milking systems: accounting for the effects of
variable intervals between milkings on milk composition.
Livest Prod Sci 73, 45–54.
Lamming GE, Bulman DC, 1976: Use of milk progesterone
radioimmunoassay in diagnosis and treatment of subfertility
in dairy cows. Br Vet J 132, 507–517.
Lamming GE, Darwash AO, 1998: The use of milk proges-
terone profiles to characterise components of subfertility in
milked dairy cows. Anim Reprod Sci 52, 175–190.
Lattec, 2007: http://www.herdnavigator.com/pages/id33.asp.
Mahalanobis PC, 1936: On the generalised distance in statis-
tics. Proc Natl Inst Sci India 12, 49–55.
Opsomer G, Coryn M, Deluyker H, De Kruif A, 1998: An
analysis of ovarian dysfunction in high yielding dairy cows
after calving based on progesterone profiles. Reprod
Domest Anim 33, 193–204.
Peters AR, Ball PJH, 1995: Reproduction in Cattle, 2nd edn.
Blackwell Science, Oxford, UK.
Prandi A, Messina M, Tondoio A, Motta M, 1999:
Correlation between reproductive efficiency, as determined
by new mathematical indexes, and the body condition score
in dairy cows. Theriogenology 52, 1251–1265.
Romano JE, Thompson JA, Forrest DW, Westhusin ME,
Tomaszweski MA, Kraemer DC, 2006: Early pregnancy
diagnosis by transrectal ultrasonography in dairy cattle.
Theriogenology 66, 1034–1041.
 2008 The Authors. Journal compilation  2008 Blackwell Verlag
121
Rossing W, Spahr SL, 1992: Automation in machine milking.
In: Bramley AJ, Dodd FH, Mein GA (eds.), Machine
Milking and Lactation. Queen City Printers Inc., Burling-
ton, pp. 311–342.
Royal MD, Darwash AO, Flint APF, Webb R, Woolliams JA,
Lamming GE, 2000: Declining fertility in dairy cattle:
changes in traditional and endocrine parameters of fertility.
Anim Sci 70, 487–501.
Royal MD, Pryce JE, Woolliams JA, Flint APF, 2002: The
genetic relationship between commencement of luteal activ-
ity and calving interval, body condition score, production,
and linear type traits in Holstein-Friesian dairy cattle. J
Dairy Sci 85, 3071–3080.
Simersky R, Swaczynova J, Morris DA, Franek M, Strnad M,
2007: Development of an ELISA-based kit for the on-farm
determination of progesterone in milk. Vet Med 52, 19–28.
Sreenan JM, Diskin MG, Morris DG, 2001: Embryo survival
rate in cattle: a major limitation to the achievement of high
fertility. In: Diskin MG (ed), British Society of Animal
Science Occasional Publication, 26 (1). British Society of
Animal Science, Edinburgh, pp. 93–104.
Van Eerdenburg FJCM, 2006: Estrus detection in dairy cattle:
how to beat the bull. Vlaams Diergen Tijds 75, 61–69.
Waldmann A, 1993: Enzyme immunoassay (EIA) for milk
progesterone using a monoclonal antibody. Anim Reprod
Sci 34, 19–30.
Waldmann A, Ropstad E, Landsverk K, Sørensen K, Sølverød
L, Dahl E, 1999: Level and distribution of progesterone in
bovine milk in relation to storage in the mammary gland.
Anim Reprod Sci 56, 79–91.
Author’s address (for correspondence): NC Friggens, Faculty of
Agricultural Sciences, University of Aarhus, Research Centre Foulum,
PO Box 50, 8830 Tjele, Denmark. E-mail: n.friggens@agrsci.dk
Conflict of interest: NC Friggens has received a research grant for this
work, part-funded by Lattec I ⁄ S; C Ridder is employed by Lattec I ⁄ S;
all remaining authors declare no conflict of interests.