Activity-based first estrus detection in grazing heifers and first-lactation dairy cows for use as novel fertility phenotypes at scale

Timing of first detected estrus of peripubertal heifers and postpartum dairy cows is an informative predictor of reproductive performance with potential to support genetic gain in fertility. Current methods of assessing reproductive status can be time consuming (e.g., requiring repeated manual observation), expensive (e.g., where laboratory analysis is required) or lack temporal resolution (e.g., blood progesterone). Additionally, on-farm collected data often suffer from quality issues. There is a growing opportunity for developing novel fertility traits for genetic selection purposes, such as age at first estrus, first estrus postpartum, and characteristics of estrus including duration and intensity, using metrics derived from wearable sensor devices. However, data capture is still often a barrier for applying sensors to extensively grazed heifers who are managed remotely, without frequent yarding, and can be outside the range of fixed receivers for extended periods. Here, we evaluated whether leg-mounted triaxial accelerometers could generate activity-based first estrus phenotypes under these constraints by relying on extended onboard data storage. In this proof-of-concept study, IceQube triaxial accelerometers were fitted to 1,880 nulliparous Holstein-Friesian (HF) and HF × Jersey crossbred heifers in 18 herds in 2019 and 1,786 primiparous HF and HF × J crossbred cows in 17 herds in 2020 (66% heifers from the previous year and 34% new animals) managed in pasture-based, seasonal calving systems. Signal-processing methods were used to identify periods of high step numbers or low lying time that predicted estrus events. Twenty-four-hour moving averages of the activity data were first calculated for each animal to remove diurnal patterns, and the herd mean was then subtracted at each time step to correct for whole herd activity. The resulting deviation was standardized for each animal to a mean of zero and a standard deviation of one, and high activity events were indicated when the standardized deviation remained above a threshold of 2.2 for at least 12 hours during a two-day period. Approaches to handling missing data, spurious detection spikes and censored results were also proposed. High activity events were detectable in both the heifer and cow step data. They were also highly correlated (F1 statistic in the range 64.1-75.9%) with estrus observations in reference datasets; these being (1) a single herd of heifers where visual observation aided by tail paint was used to record estrus, and (2) recorded insemination dates with and without subsequent positive pregnancy tests for 14 herds of first-lactation cows. The algorithm was then used to predict first estrus for each animal during the peripubertal and postpartum periods, to derive novel activity-based fertility phenotypes. We conclude that data from wearable activity sensors can be used to derive activity-based estrus phenotypes at scale, although further validation across commercial heifer herds would strengthen application in replacement animals. The approach supports the use of sensors at scale in generating useful fertility phenotypes where sensor connectivity is a challenge, although longer monitoring periods and improved data retrieval infrastructure will be needed to make data collection feasible and reduce censoring in extensively grazed heifers.

Age at first estrus and first postpartum estrus in dairy cattle are measures that could provide novel phenotypes for genetic evaluation of fertility traits. Current methods to collect measures that describe reproductive status in peripubertal heifers and early postpartum cows are labor intensive (e.g., where repeated observations are required), costly (e.g., where laboratory analyses are required), and lack temporal resolution (e.g., due to low frequency of physical visits); therefore, they offer logistical challenges and typically do not scale well for obtaining phenotypes for the large numbers of animals across many herds necessary for genetic evaluation (Palmer et al., 2010; Stephen et al., 2024). For genetic improvement, high-frequency, high-volume quantitative methods are required (e.g., Ismael et al., 2015; Heringstad and Wethal, 2023). Currently, technological innovations and subsequent on-farm adoption of wearable sensors is increasing, including for young animals (Dela Rue and Eastwood, 2023; Lamanna et al., 2025; Lee et al., 2025), and these devices may provide data that can be used for this purpose. If so, widely available, early and accurate estrus phenotypes could inform genetic selection approaches for accelerating the rate of genetic gain in fertility (Brito et al., 2025).

Emerging automated approaches to estrus detection include specialized on-cow sensors and infrared cameras (LeRoy et al., 2018; Pfeiffer et al., 2020; Nagahara et al., 2024; Borchardt et al., 2025). Wearable sensor devices such as neck-mounted collars, ear-tags, and leg-mounted accelerometers are increasingly being used on commercial farms to detect estrus events and support artificial breeding. They offer an emerging opportunity for a practical and scalable method for monitoring changes in individual animal metrics such as activity, rumination parameters and lying behaviors with a high level of detail (Reith and Hoy, 2018). However, the link between these measures (such as activity measured by step count) and estrus is indirect, requiring a predictive model to be developed. There are numerous reports on signal processing (Løvendahl and Chagunda, 2010; Duyala et al., 2014; Ismael et al., 2015) and machine learning (Yin et al., 2013; Shahriar et al., 2016; Wang et al., 2022) approaches for predicting estrus from on-cow sensor data. Signal processing is more transparent and direct when limited data streams are available, whereas machine learning is more flexible and suited to multiple data sources. In either approach, obtaining enough accurately labeled training data may be a barrier to developing novel phenotypes that can then be used as predictor traits to enhance genetic merit for fertility.

Most estrus-detection studies using wearable sensors implicitly assume frequent proximity to infrastructure, which is not often the case for extensively grazed replacement heifers, which may be managed away from the dairy platform, in large paddocks and yarded relatively infrequently. The work presented here was carried out in the context of a wider initiative to investigate using data from wearables and to develop other earlier-in-life novel measures for large-scale phenotyping of dairy cattle managed in pasture-based, seasonal calving dairy systems with the ultimate aim of improving the genetic evaluation of fertility (Steele et al., 2023; Stephen et al., 2022, 2023, 2024). To capture heifer phenotypes at scale, solutions that work under diverse heifer management scenarios are required. The specific objectives of this current study were to develop: (1) a signal processing algorithm from raw activity data in peripubertal heifers and lactating cows to predict an estrus event; (2) a method for characterizing the intensity and duration of these estrus events; and (3) a method for identifying first estrus in extensively grazed heifers and lactating cows, while accounting for missing data, spurious group level activity events, and censored estrus outcomes (see Stephen et al., 2022, 2023).

Field data collection

Data for this study were collected from 2018-born animals within a subset of herds involved in a larger study investigating novel fertility-related phenotypes in nulliparous dairy heifers and primiparous cows managed in seasonal, pasture-based systems (Steele et al., 2023; Stephen et al., 2023, 2024). Herd owners consented to participate in the main study only, as described in Steele et al. (2023), or the main study as well as this study using IceQube triaxial accelerometers (Peacock Technology Ltd, Stirling, UK; formerly IceRobotics Ltd, Edinburgh, UK). The AgResearch Animal Ethics Committee (Hamilton, NZ) approved this study and all manipulations (AE applications: 14448 and 15004). Herds were enrolled in three regions (n = 35, Waikato; n = 15, Taranaki; and n = 4, Otago) of NZ, as described by Steele et al. (2023) and Stephen et al. (2023). They were selected primarily to maximize the number and diversity of Holstein-Friesian sires represented (and resulting animal populations of Holstein-Friesian (86%) or Holstein-Friesian × Jersey crosses (13%)), the quality of existing herd records, regional location, and the willingness of the herd owners and graziers to make animals available for measurements.

The current study used 1,880 nulliparous heifers from a subset of 18 herds in 2019 (n = 11, Waikato; n = 5, Taranaki; n = 2, Otago) and 1,786 primiparous cows from a subset of 17 herds in 2020 (n = 9, Waikato; n = 6, Taranaki; n = 2, Otago; 66% heifers from 13 herds from the previous year and 34% new animals including 4 new herds). These 18 heifer herds were a subset of the 54-herd population described in Steele et al. (2023), which provides important descriptive information about the animals and their management. It should be noted, however, that some farm-level management variables such as property layout, the timing and frequency of stock movements, and walking distances were not recorded and therefore remain unknown for the herds in the current study.

The animals were fitted with IceQube devices on one hind leg to record the number of steps and lying time at 15-minute intervals. The average monitoring period for the heifers was from 18 May to 20 July 2019 (mean age 290 days at start of monitoring, mean monitoring period 64 days, standard deviation 28 days). The monitoring period was timed to coincide with the peripubertal period and ended prior to their first seasonally concentrated breeding period (either natural mating or artificial breeding) was designed to capture genetic variation in onset of puberty as described by Steele et al. (2023) and Stephen et al. (2022, 2023). Based on blood progesterone profiling data reported in Steele et al. (2023), the estimated proportions of the current study heifers reaching puberty by the beginning, midpoint, and end of the monitoring window were approximately 15%, 33%, and 50%, respectively, indicating that the monitoring period captured a meaningful proportion of pubertal transitions within the cohort.

The average monitoring period for the first-lactation cows was from 26 June 2020 to 13 January 2021 (mean age 695 days at start of monitoring, mean monitoring period 201 days, standard deviation 40 days). Monitoring covered the whole period from approximately 7 days before first calving on average, until six weeks after the end of the herd’s seasonal breeding period (typically artificial breeding for 4 to 6 weeks followed by a period of natural mating).

Animals were removed from the data set if they had less than 20 days of activity data in the relevant period; this resulted in 202 heifers and 17 cows being excluded from the datasets. Excluded animals occurred in 6 cow herds and 17 heifer herds, and had age and calving date ranges similar to non-excluded animals.

The research mob consisted of 129 heifers of which 75 heifers served as a ‘heifer activity-reference’ group. They were managed by the Bioeconomy Science Institute (formerly AgResearch) at Ruakura (Waikato, New Zealand, 37.7747°S, 175.3103°E) and observed by DairyNZ technical staff for behavioral signs of estrus three times weekly on a Monday-Wednesday-Friday basis from 15 April to 2 September 2019. This observational assessment, performed to identify approximate estrus date, was aided with the tail-painting technique, where tail-paint wear for every heifer was scored on a 3-point scale (1 = mostly removed; 3 = untouched) at each visit. Inferred estrus event periods were denoted as having “Strong” (n = 240) or “Weak” (n = 110) likelihood of estrus, based on the degree of tail paint/mounting, where “Weak” = no mounting observed but tail paint disturbed, and “Strong” = mounting observed and tail paint removed. All reference heifers had between one and seven events recorded. There was no such behavioral data available from the other 17 heifer herds.

For 14 of the 17 cow herds, artificial insemination (AI) dates and conception dates confirmed via fetal-aged pregnancy diagnoses served as benchmarking data in model development of the IceQube data. Conception dates were determined by experienced personnel using transrectal ultrasonography to age the fetus to a recorded AI date. Subsequent pregnancy data were supplied by herd record providers (LIC, Hamilton, NZ; CRV, Hamilton, NZ). The AI dates associated with pregnancy success were labeled as “Gold” events (n = 878), whereas other AI dates were labeled as “Silver” events (n = 973). These labels, Gold and Silver, should not be interpreted to convey any differences in intensity or duration of estrus. Signal calibration was determined using these AI events.

The IceQube data were uploaded to the IceRobotics Ltd. central database and subsequently extracted to a Snowflake database (Snowflake Inc., Bozeman, MT, USA). Data were analyzed using the “tidyverse” package (Wickham et al., 2019) in R (R Core Team, 2023). Plots were produced using the “ggplot2” package (Wickham, 2024). Time stamps in the raw device data were converted to local time and rounded to the nearest 15 minutes. Duplicate time steps and animal IDs were removed. Periods of missing data were noted for some devices and for the whole day of 5 October 2020 (UTC) for all devices and herds (due to a data transfer issue) as provided by the manufacturer.

A visualization of daily variations in raw step count for an example herd of 49 first-lactation cows is presented in Figure 1. Devices were attached to cows in mid-June 2020, prior to calving, which coincides with a period of low walking activity. After calving (red triangles), step counts increase with twice-daily walking from the grazing paddocks to the milking parlor. Gaps indicate missing data (more than 1 hour missing in a day) and dark periods in December indicate sensor data after removal from the cows; that is, prior to downloading data from individual devices. A key objective of this work was to automatically identify the peaks in activity associated with estrus for individual animals, sometimes visible as orange to dark red dots in Figure 1. The pattern for raw lying time data appears similar, albeit with the green and blue regions reversed (data not shown).

Signal processing

We adapted the approach of Løvendahl and Chagunda (2010) and Ismael et al. (2015) to develop a signal processing algorithm based on the activity time-series data. Unlike the voluntarily milked indoor cows in those prior studies, lactating cows in a grazed system have regular high activity periods walking to and from the milking parlor over varying distances, as well as having more marked diurnal behavior patterns due to timing of allocation of fresh pasture (typically either twice or once daily following a milking event). Further, nulliparous heifers in a grazing system are often grazed on undulating to steeper terrain, may exhibit more playful active behaviors and have less frequent visits to yards, e.g., for monitoring live weight and receiving animal health treatments. Data processing therefore was required to isolate signals due to behavior associated with estrus among a highly active animal dataset.

The raw IceQube activity data (a t,j steps or lying time at time t for cow j recorded at 15-minute intervals) were processed using standard signal processing techniques (Ogasawara et al., 2010) to identify periods of high activity of individual animals relative to their herd mates, as follows. First, a simple right-aligned moving average for the previous 24 hours was calculated (r t,j). This smooths the individual cow signal and corrects for any diurnal behavior patterns characteristic of individual cows. Note, however, that using such a right-aligned 24-hour moving average widens and delays any event signal. This event signal is ‘triggered’ by the onset of increased activity and the event date is established accordingly. In practice, the moving average method has minimal effect on the timing of this trigger, although the peak is delayed by 12 hours, and the duration is extended by 24 hours. Moving averages for 24-hour periods with more than one hour (4 time steps) of missing data were removed.

Second, based on inspection of the normal range of the time series data, steps data with r t,j outside the range [10, 200] steps per 15-minutes and lying time data with r t,j outside the range [300, 895] seconds per 15-minutes were considered outliers, and removed. Based on these empirical criteria, 3.4 million of the 39.5 million rows of data were removed (8.7%, range 0.5% to 20.9% across farms).

Third, the herd mean r t,j value at each point in time was subtracted from the data for individual animals to calculate a deviation (d t,j). This step corrects for activity associated with herd management, such as moving to a new paddock or milking. Ismael et al. (2015) used a ratio instead of a difference, although this risks error distortion when the denominator is small. If fewer than 6 cows were available to calculate the herd mean value, then data for that period were removed. This was based on empirical observations that herd standard deviations were unstable when less than 6 cows were available.

Fourth, the d t,j series was standardized for each cow using her mean and standard deviation over the data set to produce a standardized deviation index (z t,j). Since the d t,j series includes both estrus and non-estrus periods, this scaling was affected by the animal’s strength of estrus. However, estrus represents a relatively small proportion of an animal’s overall signal, so this effect is likely to be small.

The steps for calculating z t,j at time t for cow j from the raw activity data a t,j were therefore:

Start with a t,j = activity in time step t for cow j (e.g., steps, lying time).

Calculate r t,j = right-aligned 24-hour rolling average = m e a n u > t − 24 h u = t (a u, j).

Consider as missing, values of r t,j based on less than 23 hours of data, or with extreme values.

Calculate d t,j = deviation from herd mean = r t, j − m e a n k = 1 n j (r t, k).

Consider as missing, values of d t,j based on data from fewer than 6 cows.

Calculate z t,j = standardized deviation = d t, j − m e a n u = 1 n t (d u, j) s d u = 1 n t (d u, j).

Figure 2 shows an example for an individual animal as a nulliparous heifer (2019) and a primiparous lactating cow (2020). Estrus signals calculated using this method are depicted, based on steps (blue) and non-lying time (red), respectively. Both metrics result in similar patterns in this case, due to both sensor metrics being calculated from the same underlying 3-dimensional accelerometer data. Similar patterns have been observed in tropical cattle (Duyala et al., 2014). Further analysis (presented below) showed, however, that lying time is a less useful metric for lactating cows compared with heifers, likely due to cows being brought in for milking twice a day having less opportunity for lying and a greater drive to graze to meet the greater feed demands of lactation.

Estrus detection parameters

Automatically detecting an estrus event from the processed activity data required setting several parameters, particularly the threshold above which “high activity” is deemed to occur (T) and the duration above this threshold that indicates an estrus event (D). A minimum number of data points available (N) is also required. These parameters were estimated to match the manual observations of the heifer reference herd and cow AI dates, as described below.

The following decision logic was used to identify whether an estrus event was predicted on a particular 24-hour day, based on the 15-minute smoothed deviations of the step or the non-lying data for that day and the following day (e.g., Figure 2), since estrus duration was expected to be less than 48 hours:

If the deviation is above T for at least D hours of the current day, then an estrus event is predicted;

Else, if the number of data points for the current day is less than N, then the prediction is flagged as unknown;

Else, if the deviation remains below T during the current day, then no estrus event is predicted;

Else, if the deviation is above T for at least D hours of the current and following days combined, then an estrus event is predicted;

Else, if the number of data points for the current day and the following day combined is less than 2N, then the prediction is flagged as unknown;

Else, no estrus event is predicted.

If estrus events are predicted on consecutive days, only the first day is marked as a predicted estrus event and subsequent days of the same event are not marked. Predicted estrus events during the first 2 days after device fitting on heifers were excluded, and predicted estrus events during the first 16 days after calving were excluded for cows (to avoid spurious activity associated with the peri-calving event; Crowe, 2008; Häggman et al., 2019; Borchardt et al., 2021).

A range of possible values for the parameters T and D were tested to see what threshold values gave the most accurate predictions when compared with the manually observed cow (AI dates) and heifer estrus events from the reference herd. From Yin et al. (2013), T was expected to have a value around 2. Based on the typical duration of estrus in HF animals under a grazed pasture system (Reed et al., 2022, 2024), D, the minimum duration, would be expected to be less than around 11.3-12.5 hours for primiparous cows and 14.1-15.1 hours for heifers. However, as previously mentioned, the simple moving average smoothing of the data spreads the signals over a longer time period (Figure 2) so values of D could overestimate the physiological minimum of estrus duration.

A limitation in calibrating the approach was that the manual assessment data were not precise. For the cow AI data, for example, the observed estrus probably began on the previous day or earlier on the same day as the recorded AI date. For the heifer tail paint/mounting data, timing of estrus was reported on the basis of three observations per week. In order to align the automatically predicted estrus events with the manual observations, once the signal processing had been completed and estrus event dates predicted from the steps or lying activity data, the timings of the manually observed event dates were adjusted to coincide with predicted event dates that occurred within the possible date range provided for each observed estrus (± 1 day for cow events, 1–10 day window for heifer events). The predictive performance metrics described below are subject to this adjustment step. Following this adjustment, the number of true positive (TP), false positive (FP) and false negative (FN) predictions were then counted for each herd, and summed across herds, for estrus events predicted from the steps or lying data, separately. Three metrics were used to evaluate predictive accuracy: recall (Rec, the proportion of observed events that were correctly predicted, also known as sensitivity), Precision (Pre, the proportion of predictions that corresponded with observed events, also known as 100% minus the error rate) and the F1 statistic (the harmonic mean of Rec and Pre) (Saito and Rehmsmeier, 2015; Reith and Hoy, 2018). These metrics are calculated as:

In order to estimate the predictive performance relative to the Strong/Weak observations, or the Gold/Silver observations, separately, the false positives had to be assigned to Strong/Weak or Gold/Silver, respectively. This was done by assuming that false positives occurred in proportion to the number of Strong/Weak observations or Gold/Silver observations in the dataset. Analysis of predictive performance was carried out over a grid of T and D values (T = 0.0, 0.2, 0.4, 0.6 … 5.0; D = 3.0, 4.5, 6.0, 7.5 … 21.0 hours) to identify the optimal values that maximized F1. These optimal parameter values were first determined for each herd separately, and then for the pooled observations. This provided a form of cross-validation that the parameters values derived from pooled calibration performed acceptably across all herds in the study. The parameter values from the pooled calibration were then used to predict estrus events in the heifer and cow datasets.

Outlier removal

In cases where there were unusually high numbers of predicted estrus events on the same day (i.e., “spikes” in the numbers of predicted estrus events), additional outlier removal was applied (further investigation suggested these spikes could have occurred when the herd was not being managed as a single mob which may reflect higher activity due to one-off management procedures such as weighing or regrouping of mobs for mating). The occurrence of spikes was assessed relative to the frequency of estrus detections at the time; for example, smaller spikes in estrus detection were more notable early in the monitoring period when few cows were cycling, while similar percentages of cows in estrus later in mating would be considered unremarkable. The percentage of animals in estrus on each day was therefore compared with the smoothed pattern of daily estrus predictions for that herd (calculated using a gam model assuming a “cs” cubic regression spline basis and a “scat” residual distribution, to be less sensitive to outliers) to determine when notable spikes may be occurring. These data were aggregated across all herds to develop a common threshold equation for spike detection expressed as a linear function of the smoothed average percentage of estrus on each day of the season for that herd, and used to identify the days with abnormally high numbers of event ‘spikes’ which exceeded this threshold. Predicted estrus events on these days were assumed to be spurious and these days were ignored for all cows. The results of this procedure are detailed in the Results.

Inter-estrus intervals and censoring

For heifers or cows that had more than one estrus event predicted, the interval between consecutive predicted events was analyzed to determine the range of typical intervals. This was done by removing extreme outliers using repeated application of the Tukey criterion (Tukey, 1977); values outside the outer fences, i.e., below Q1 – 3 × IQR or above Q3 + 3 × IQR, where Q1 and Q3 are the first and third quartiles and IQR is the interquartile range Q3 – Q1, were considered extreme outliers, and were removed. The median and outer fence values for valid inter-estrus intervals were then determined. We accepted inter-estrus intervals between 13 and 27 days to be valid for our purpose, which was to support phenotype construction, and not define all possible cycle lengths. Consecutive predictions spaced within this range (i.e., 13–27 days) were considered as genuine estrus cycle patterns.

Onset of the earliest predicted inter-estrus cycle event for each heifer and cow was then determined, if available, otherwise the earliest single predicted estrus event was recorded. Left- and right-censoring was applied as described in Table 1. Left-censored predictions provide only an upper bound to the time of the first estrus event, whereas right-censored predictions provide only a lower bound to the time of the first event. The earliest predicted estrus event was left-censored when data were available for less than half of the time period when a previous estrus event may have occurred (i.e., 8+ days missing data in the period 27–13 days prior to the earliest event). When no events were detected during the monitoring period, the predicted first estrus date was considered right-censored (Stephen et al., 2022). Note that this could include some animals in the cow dataset, for example, whose first estrus event was not detected but who became pregnant during the data recording period.

Combining the predicted age at first estrus determined from the activity data for each herd of heifers, or the predicted days to first estrus after calving for each herd of cows, allowed the estrus distribution to be plotted for each herd in each year. Due to the presence of left- and right-censored results, a cumulative distribution function (CDF) plot was used to compare results between herds, as this allows left- and right-censored values to be represented naturally. For each age class, an average CDF across all herds was also generated, based on the pooled mean and standard deviation inferred from the censored data using the “survreg” method in the “survival” package in R (Therneau, 2024).

The strength and duration of estrus events were also of interest. For each predicted estrus event, the maximum amplitude (of d t,j, steps per 15-min above herd mean), duration (hours that z t,j was above the threshold T) and total area under the curve (integral of d t,j while z t,j was above the threshold T) were calculated. These metrics were compared for all herds. Only the results for the step data are presented here; lying time results are not shown.

Estrus detection

A comparison of AI events and model predicted estrus events for the cow reference herd is presented in Figure 3. Calving dates are shown as red triangles, and the assumed earliest possible days of estrus, 16 days after calving, are indicated as red dots. For the activity data, dates with N ≥ 48 time-steps (12 hours) of processed (smoothed, relative, standardized) steps data are shown as gray dots. Predicted events (using parameters T = 2.2, D = 12 hours in this case) are shown as blue dots. Days identified as having an abnormally high percentage of predicted events (i.e., spikes or outliers, discussed below) are shown as vertical lines of red crosses: these were not removed when calibrating the event detection parameters, but were removed for phenotype analysis. Predicted events identified as being part of inferred cycles are highlighted with blue circles. Gold and Silver AI benchmark dates are shown as gold and silver diamonds, respectively.

To assess predictive performance, manually observed events were adjusted to coincide with the predicted events that occurred within their accuracy windows (± 1 day for cow AI dates). Recall, precision and F1 statistics were then calculated as described above. Looping through a range of T parameters for each herd and for steps and lying time (while holding D constant at 12 hours in the first instance) allowed model prediction performance to be quantified. Figure 4 shows the precision and recall results across different critical deviation values (with minimum duration D set at 12 hours) relative to the 14 cow herds (top) and 1 heifer herd (bottom) that had manual estrus observations, for steps (left) and lying time (right). The solid lines summarize the results for all events, while the dashed and dotted lines show the results against the Strong and Weak observations for the heifers, or the Gold and Silver observations for the cows, respectively. For cows, the Gold events (AI dates with pregnancy success confirmed) were easier to predict than the Silver events (AI events that failed to result in pregnancy success). The recall of Gold event prediction also peaked at higher values of T compared with the Silver events, for both steps and lying. For heifers, recall and precision of Strong events (observed mounting) was more modest compared with the cow events, and recall of Weak events (observed disturbance of tail paint) was quite poor. Similar to the cows, recall of Strong events for the heifers peaked at higher values of T. Optimal values of T for cows were 2.4 for Gold events based on steps and 2.0 based on lying time. For heifers, the optimal values of T were 2.0 for Strong events based on steps and 2.2 based on lying time. This consistency supports the hypothesis that T values around 2 are optimal for detection of true events, aligning with Yin et al. (2013), who also used a threshold of 2.

To optimize D, the F1 predictive performance metric was calculated across a range of T and D values for the step and lying time data of the heifer and cow herds, both individually and combined (Figure 5). As expected, using a higher threshold (T) parameter worked best with a lower duration (D) parameter, due to the peaked shape of the typical event signals (as in Figure 2). Optimal values of T and D for individual herds were quite variable (shown as circles in Figure 5), but due to the flat ridged shape of the F1 surface, using the optimum values for the pooled herds (T = 2.2, D = 12.0, shown as squares in Figure 5) gave F1 values at least 93% of the individual herd optimum F1 values in all but 2 cases (as indicated by the size of the circles in Figure 5).

Using these parameters with the step data, the overall predictive performance for the cows was recall 80.9%, precision 71.5% and F1 75.9%, and for the heifers was recall 62.7%, precision 72.8% and F1 67.3%. All performance metrics were lower when the lying time data were used.

Outlier removal

For several herds, there were days where the percentage of animals predicted to be in estrus was far higher than the average daily percentage. In Figure 3, these “spikes” are indicated by vertical lines of red crosses. For all herds, the daily percentage of animals predicted to be in estrus was compared with the smoothed average percentage of animals predicted to be in estrus (based on steps or lying data). For example, Supplementary Figure 1 shows this comparison for a herd of heifers in 2019 and in the following year as cows. Spikes in detections tended to occur early in the heifer data, whereas spikes in detections in the cow data tended to occur in mid-mating or even after mating was completed.

To systematically identify these outliers, a threshold was proposed where the percentage of estrus per day exceeded the smoothed average percentage of estrus per day × 2.4 + 5% (Supplementary Figure 2). This threshold was used to identify the days with abnormally high numbers of event ‘spikes’ (two such spikes are shown for example in Figure 3 as vertical lines of red crosses). In total, 24 days were identified as spikes in the cow step data, out of a total of 3658 days (0.6%), while 20 days were identified as spikes in the heifer step data, out of a total of 1349 days (1.4%). This resulted in the removal of 288 predicted cow events based on steps, out of a total of 5026 (5.7%), and the removal of 236 heifer events based on steps, out of the total of 2804 (8.4%). For the lying data, 38 days were identified as spikes in the cow data, out of a total of 3639 days (1.0%), while 36 days were identified as spikes in the heifer data, out of a total of 1333 days (2.7%). This resulted in the removal of 426 predicted cow events based on lying, out of a total of 4703 (9.1%), and the removal of 510 heifer events based on lying, out of the total of 2367 (21.5%).

Inter-estrus intervals and censoring

Having removed estrus events identified as erroneous spikes, the inter-estrus intervals were analyzed. Figure 6 shows the distribution of these inter-estrus intervals for the heifer and cow data. By repeated application of the Tukey criterion for extreme outliers, median and outer fence values were determined. These suggested 13–27 days as a suitable range for valid inter-estrus intervals for both cows and heifers, with 20 days as the median. These ranges align with field observations, as discussed in more detail in the Discussion.

In this study, the 13–27 day inter-estrus interval range was assumed acceptable to support labeling of estrus events that were considered genuine (as marked with blue circles in Figure 3). The first such estrus event for each animal was used to predict her phenotype; that is, her age at first detected estrus (heifers) or days to first estrus after calving (cows). If more than half of data were missing in the period prior to the first cycle or event (i.e., the period where a previous, undetectable estrus may have occurred), the first predicted estrus event was considered to be left-censored (Table 1). If only a single event was detected for an animal, this was taken as the phenotype, whereas if no events were detected for an animal, then her age at the final day of recording was used to predict her phenotype, which was considered to be a right-censored prediction.

Following this approach, phenotypes were determined for each heifer in 2019 (that is, age at first detected estrus) and for each cow in 2020 (that is, days to first estrus after calving).

Predicted age at first estrus in heifers

The distribution of heifer age at first detected estrus (as a proxy for age at puberty) was then plotted as a cumulative distribution function (CDF) for each herd (Figure 7) following Macdonald et al. (2005) and Meier et al. (2021). These curves show the cumulative proportion of heifers that had a detected estrus event by a certain time, and the proportion of heifers who may have had a detected estrus either before the sensors were applied or after they were removed, indicated by the missing y-values at the left- and right-hand ends of the curves, respectively. The large left- and right- censored portions of some of the CDF curves suggest that the age of first detected estrus may not have been accurately captured for animals in some herds.

Based on the 1,117/1,678 (67%) of the heifers with non-censored results, the median age at first detected estrus in heifers was 330 days (range 243-400). Taking into account censored results using survival analysis, the mean was estimated to be 334 days with a standard deviation of 36 days; a normal distribution with this mean and standard deviation is included as a cumulative probability distribution in Figure 7 for reference. The large proportion of censored results in the heifer data makes it difficult to compare the individual herds with certainty, but it appears that some herds reached first estrus much earlier than others.

The relationship between estrus duration (Hours above threshold), maximum estrus activity (Max, steps per 15-min above herd mean) and the integral of activity above the detection threshold (Total, steps above threshold) were compared (Figure 8). Estrus duration ranged (5 th to 95 th percentile) from 15.5 to 32.0 hours above threshold with a median of 24.0 hours. Maximum estrus activity ranged from 11 to 112 steps per 15-min above the herd mean with a median of 53 steps. Total event steps ranged from 75 to 4,769 steps above threshold with a median of 1,629. Since the shape of estrus activity tends to follow a “hump” shape (Figure 2) the strength, duration and total activity are correlated. The correlation between maximum estrus activity and duration was 0.34.

Predicted days to first estrus after calving in cows

Similarly, the first estrus predicted after calving was recorded for each cow. The distribution of days from calving to first estrus was then plotted for each herd as a cumulative probability distribution (Figure 9). Compared with the heifers, there was a generally lower proportion of left-censored results.

Based on the 1,566/1,771 (88%) of the cows with non-censored results, the number of days after calving to first estrus had a median of 45 days (range 16–188 days). Taking into account censored results using survival analysis, the mean was estimated to be 58 days with a standard deviation of 39 days; a normal distribution with this mean and standard deviation is included in Figure 9 as a dotted line for reference. The herd results have a steep initial shape, however, suggesting that the true distribution of days to first estrus after calving is right skewed within a herd.

The relationship between estrus duration (Hours), maximum estrus activity (Max), and the integral of activity above the detection threshold (Total) were compared (Figure 10). Estrus duration ranged (5 th to 95 th percentile) from 17.5 to 31.0 hours above threshold with a median of 24.8 hours, slightly longer than the heifer estrus duration. Maximum estrus activity ranged from 23 to 96 steps per 15-min above the herd mean with a median of 51 steps. Total event steps ranged from 324 to 4773 steps above threshold with a median of 1923. The correlation between maximum estrus activity and duration was 0.33.

In this proof-of-concept study, we present an improved signal processing method that can be applied to activity data collected by on-animal sensors to identify and characterize estrus events in peripubertal heifers and first-lactation cows resuming reproductive cycling after calving. This method accounts for changes in animal activity relative to other animals under the same management, the time of day, and the individual’s typical behavior characteristics. Overall, our approach achieved an acceptable level of performance (Recall 59.1-80.9%, Precision 67.1-72.8%, F1 64.1-75.9%, using the step data) as a scalable method for providing novel phenotypes, such as age at first estrus in heifers and time to first postpartum estrus in cows, for genetic evaluations of fertility in dairy cattle. For genetic evaluation purposes, we need not have the level of precision often required in making correct management decisions (e.g., submitting an individual cow to AI), but instead require a consistently applied method providing a useful genetic signal from large numbers of animals managed across many herds. The current approach was designed to minimize herd- or management-specific biases. The validation methodology used in the current study serves for a ‘proof-of-concept’ research objective based on a limited amount of observational data. We recommend that further, more robust, validation is conducted before the proposed methodology is implemented in practice.

Nevertheless, our performance results for the first-lactation cows compare favorably with those reported for commercial devices (Dela Rue et al., 2013; Chanvallon et al., 2014; LeRoy et al., 2018; Fauvel et al., 2019; Borchardt et al., 2025). Reith and Hoy (2018) cited several studies indicating that, based on activity data alone, recall can range from 37 to 90% and precision from 55 to 83%. Performance is variable, as are the devices available and the scenarios under which performance is assessed. Borchardt et al. (2025), for example, estimated recall and precision averaging 70.3% and 88.1%, respectively (F1 78.2%) for several types of devices used on cows during the voluntary waiting period, which included the period where cows resume having postpartum ovulatory cycles and displays of behavioral estrus. The benchmark in this previous case was ovulatory patterns based on progesterone measures. Our results based on step data for first-lactation cows had an average recall of 80.9%, precision of 71.5% and F1 of 75.9%. The higher recall value in the current study is likely explained by training against AI dates that followed detection of estrus by the herd managers, rather than ovulatory patterns in early postpartum that may or may not be accompanied with overt estrus. Kamphuis et al. (2012) reported on a collar mounted activity system that achieved a recall of 76.9% and a precision of 82.4% (F1 79.6%) when benchmarked against ovulatory cycles determined by milk progesterone monitoring during a seasonal breeding period in grazed dairy cows. This is a similar result under similar cow management conditions to the present study. Taken together, the performance statistics generated from the first-lactation cows in the current study are consistent with the literature and compare favorably, given the ranges reported. This provides us with confidence that the conceptual approach taken to process raw step activity data into alerted estrus events is sufficiently robust to further develop for genetic evaluation purposes. The next steps could involve investigating data metrics from other wearable sensors, obtaining a larger data set of alternative reference measures for validation purposes, exploring the effect of parity, and investigating other statistical methods for their impact on genetic evaluations of fertility. Estimation of genetic parameters from wearable-derived phenotypes for first estrus and their correlations with established fertility traits such as timing of calving or pregnancy, along with validation against alternative visual, endocrine, or ovarian measures of estrus events or ovulatory patterns, should also be considered as future steps toward implementation of these novel phenotypes within breeding programs.

The performance values for detecting age of first estrus events in peripubertal heifers were slightly less than for lactating dairy cows; this may be due to heifers displaying generally weaker activity during first estrus compared with subsequent cycles (Reed et al., 2024), and was limited by only having observed estrus records for one of the monitored herds. Estimating age at first estrus in the heifers was also challenged by the short duration of sensor use in this study, which was limited by the maximum on-board storage capability of the devices available at the time and the practicality and challenges of visiting multiple herds across different regions. Accelerometers were fitted at 270 days of age on average (range 241–306 across herds) leading to 11.8% of heifers receiving left-censored predictions of first estrus on average (range 0.0-30.6%). The devices were also removed early, at 378 days of age on average (range 361–427 across herds) resulting in 24.1% of heifers not having an estrus detection prior to removal (i.e., right-censored, range 1.9-56.8%). This relatively large proportion of censored predictions also compromised the reliability of the predicted cumulative distribution functions for the herds. While this creates uncertainty around the estimates for particular animals or herds, mean results at scale may still be valid for the purpose of genetic comparison of time-dependent fertility traits, as shown by Stephen et al. (2022, 2023, 2024). For example, using blood progesterone data collected across three visits at about 30-day intervals to the 54 herds (n = approximately 5,000 heifers) enrolled in the main study, Stephen et al. (2023) demonstrated that suitable age at puberty phenotypes for ranking sires can be derived from as little as one blood progesterone data point per offspring. Moreover, issues with left- and right-censoring of age at first estrus phenotypes are likely to be reduced in the future as technological innovations (e.g., use of direct-to-satellite linked receivers, longer-range radio receivers, or longer on-board storage) and widespread adoption enable collection of wearable sensor data over longer time periods in dairy heifers managed under extensive grazing conditions.

There appeared to be considerable variation in the predicted age at first estrus between heifers and between the median of predicted age at first estrus between heifer herds. Variations in the farm environments and animal management could possibly have contributed, although a study limitation is not having sufficient information to account for variation in these farm-specific factors. A mitigating point, however, is that the signal processing approach does account for the animal’s own activity routines in relation to the herd average routine. This feature should minimize any bias created by differences in herd (e.g., breed, condition) and farm features (e.g., walking distances).

The predicted mean of 334 days and standard deviation of 36 days are earlier but similarly spread compared with estimates reported in other studies. In the wider study involving 54 heifer herds, Steele et al. (2023) reported the mean age at puberty as 352 days, with a standard deviation of 34.9 days (median 359 days and range 251 to 429 days). There are several possible explanations for the different results, including the censoring approach (Stephen et al., 2023), the difference in the method for capturing puberty (progesterone based vs. behavioral based), and the higher resolution data (daily activity) compared with the low-resolution blood progesterone data (3 visits at 28-day intervals). It is also possible that the 18 heifer herds fitted with IceQube devices were better managed for growth rates relative to the general 54-herd population. Later age at puberty averages were also reported by Macdonald et al. (2005) for HF dairy heifers in New Zealand; 356 days and 380 days, across two cohorts of animals in consecutive years, with the difference attributed to restricted feeding in the second year. Meier et al. (2021) likewise reported median age at puberty of 358 and 385 days for New Zealand HF heifers with high and low genetic merit for fertility, respectively, using a collar-mounted heat detection system. Results from these three studies are closely aligned and suggest the age at first detected estrus values derived in the current study are earlier than routinely expected, perhaps due to trial protocol (accelerometers fitted too late) or analysis method (identification logic for first estrus activity). In general, however, the current results demonstrate sensible age at first estrus values.

Days to first detected estrus after calving was more reliably assessed than age at first detected estrus in peripubertal heifers, most likely due to the longer period over which sensors were used, the effect of calving date in constraining the possibility of left-censored results, and the lower proportion of right-censored results (animals without a detected estrus). The distribution of days to first estrus after calving was more consistent among cows and across herds. Visually, we could identify obvious differences in the rate of cows having first estrus between herds. The predicted distribution of days to first estrus (mean 58 days, standard deviation 39 days) calculated in the current study is consistent with observations of primiparous HF cows managed within pasture-based systems (McDougall et al., 1995; Burke et al., 1995). Primiparous HF cows are widely reported to take longer to resume cycling postpartum than mature cows or cows with Jersey breeding, especially under pasture-based management systems (Rhodes et al., 2003). While the mean interval to first detected estrus (e.g., 58 days) is long compared with most reports in the literature, it is consistent with that expected of primiparous HF cows on pasture.

An analysis of inter-estrus intervals was conducted as a further sensibility check of the predicted estrus events, with an expectation that these intervals would be consistent with typical reproductive cycle patterns. Inter-estrus intervals estimated from the activity data were 20 ± 7 days for both the heifers and the cows, consistent with mean values of 21.0 to 25.1 days derived by Cummins et al. (2012) and mean intervals of 21.1 reported in Lewis and Newman (1984). In addition to mean interval being consistent with the literature, the distribution pattern resembles that expected in large-scale industry data. Distinctive peaks are reported to be evident at about days 10, 21 and 42 in large-scale datasets (Macmillan, 1970; Greenham et al., 2019). The pattern of inter-estrus intervals of the current study are consistent with those seen in practice, which further supports the sensor method for accurate signal capture.

Estrus event characteristics such as the intensity and duration of heat signals may provide novel traits for fertility (Brito et al., 2025). In the present study, the duration of the processed activity signal above threshold was found to be greater than 15.5 hours for nulliparous heifers for most estrus events and greater than 17.5 hours for first-lactation cows. This is longer than the averages of 11.3-12.5 hours for primiparous cows and 14.1-15.1 hours for prepubertal heifers in a seasonal calving, pasture-based dairy system, recently observed by Reed et al. (2022, 2024), and is at the upper end of the 5.8-16.1 hour durations reviewed by Reith and Hoy (2018) for housed animals. A likely contributing factor to this difference is the 24-hour moving average applied in the current study, which tends to lengthen the duration of the signal. Since the objective in the current study was to provide a reliable fertility phenotype to compare animals for genetic selection purposes rather than to measure estrus length per se, a small degree of systematic bias is acceptable. Nevertheless, opportunities to refine the methodology to more accurately reflect the true duration of estrus should be examined to optimize the suitability of these sensor-derived traits as predictors of genetic merit of fertility.

Increased use of wearable activity monitors on dairy heifers and cows provides an opportunity for timely estrus detection and fertility phenotyping at scale. Applying simple signal processing methods to step and lying time data from accelerometers allowed us to quantify the timing, strength, duration and total activity of first detected estrus in peripubertal heifers and first-lactation cows after calving. These traits are biologically related to reproductive development and postpartum cyclicity and may be useful as fertility-related phenotypes. Although the short period of sensor use resulted in a large proportion of missing or censored results for heifers, the method used in this proof-of-concept study was able to determine differences between animals and between herds both for heifers and first-lactation cows. The degree to which these differences can be ascribed to sire genetics versus feeding or other effects has yet to be explored. Nevertheless, this approach appears feasible for estrus detection and fertility phenotyping at scale to support genetic gain. Further investigations are warranted to assess correlations with established fertility traits using the current datasets. Additionally, future research should explore the use of fertility phenotypes derived from various wearable sensors. As technological advancements increasingly provide more options, these sensor-derived phenotypes can help to overcome the constraints to obtaining robust measures suitable for ranking sires and genetic evaluations of fertility, especially in animals managed under grazing conditions.