Over recent years animal welfare has become a major research area in animal husbandry. This increasing interest has been driven both by ethical concerns and by the increasing attention of consumers on product quality rather than quantity (Thornton, Reference Thornton2010). One of the most accepted definitions of the animal welfare concept is the ‘five freedoms’, formalised in July 1979 in a report by the UK's Farm Animal Welfare Advisory Committee. A review by Webster (Reference Webster2001) is available, where these ‘freedoms’ are listed and explained. Nevertheless, being a multidimensional factor (Fraser, Reference Fraser1995), animal welfare assessment is not straightforward and it is dependent on different human cultures, traditions and religious faiths (Szűcs et al. Reference Szűcs, Geers, Jezierski, Sossidou and Broom2012).
Rumination is described as the process of regurgitation, re-mastication, salivation, and swallowing of feed to reduce the particle size and enhance fibre digestion (Erina et al. Reference Erina, Cziszter, Acatincăi, Baul, Tripon, Gavojdian, Răducan and Buzamăt2013). Rumination Time (RT, i.e., the number of minutes spent by a cow during a determined time interval) has been associated with rumen welfare, since it increases the production of saliva, which acts as a buffer for the ruminal pH (Beauchemin, Reference Beauchemin1991). The development in the early 2000s of automatic systems able to record and store a large amount of different parameters related to milk yield and cow activity, including RT, increased the possibility to investigate changes in RT and its relationships with other animal-related factors. Some studies have shown that a RT decrease might be an indicator of unfavourable psychological (acute stress: Herskin et al. Reference Herskin, Munksgaard and Ladewig2004; anxiety: Bristow & Holmes, Reference Bristow and Holmes2007) and pathological (hypocalcaemia: Hansen et al. Reference Hansen, Nørgaard, Pedersen, Jørgensen, Mellau and Enemark2003) conditions. More recently, RT has been further investigated to assess its relationship with the physiological changes linked with calving and oestrus events. Clark et al. (Reference Clark, Lyons, Millapan, Talukder, Cronin, Kerrisk and Garcia2015) correlated RT and activity time, concluding that there was a distinct decline in the duration of rumination pre-partum, which could be successfully used to predict the cows’ day of calving. Dolecheck et al. (Reference Dolecheck, Silvia, Heersche, Chang, Ray, Stone, Wadsworth and Bewley2015) described the oestrus-related changes in parameters automatically recorded by different commercial systems and assessed the potential use of this data collecting technology for oestrus detection. The relationship between RT and diseases has not been fully investigated yet. Some recent studies showed that common dairy farm diseases significantly decrease the RT (Van Hertem et al. Reference Van Hertem, Maltz, Antler, Romanini, Viazzi, Bahr, Schlageter-Tello, Lokhorst, Berckmans and Halachmi2013; Liboreiro et al. Reference Liboreiro, Machado, Silva, Maturana, Nishimura, Brandão, Endres and Chebel2015; Talukder et al. Reference Talukder, Kerrisk, Clark, Garcia and Celi2015). Stangaferro et al. (Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordano2016a, Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordanob, Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordanoc) demonstrated that metabolic and digestive disorders, mastitis, and metritis have a negative effect on RT and could be predicted by analysing patterns in RT changes.
The hypothesis tested in this study is that, by predictive modelling, a trait recorded by automatic systems (e.g., RT) could be used as predictive tools for incoming diseases. Furthermore, the aim of this study was also to describe changes in RT in the days before the recording of different diseases.
Material & methods
Data collection
The animals monitored in this study were 259 Italian Holstein cows reared in a commercial farm located in Mantua province, Lombardy (Northern Italy). All the animals were fed total mixed ration (TMR), milked twice a day and grouped in pens (lactating, pre-calving, and infirmary). RT data were recorded using the Heatime HR system (SCR Engineers Ltd., Netanya, Israel) from the 24th of September 2014 to the 6th of October 2015, for a total of 377 consecutive days. This system is composed of a neck collar with a tag containing a microphone to monitor rumination and an accelerometer to quantify activity (as validated by Schirmann et al. Reference Schirmann, von Keyserlingk, Weary, Veira and Heuwieser2009). The raw data are then processed and summarised as 2-h intervals by the herd management software DataFlow II (SCR Engineers Ltd.), where all the information regarding each single animal (e.g., ID number, age, parity) is recorded, and then downloaded in a spreadsheet file.
The list of diseases was obtained from the farm management software, where they were recorded soon after the veterinary diagnosis, both during routine or requested visits to the farm. Their incidence is reported in Table 1. All of the recorded diseases were used in this analysis, regardless of their known effect or association with RT changes. Excluding mastitis, other diseases recorded in the software were grouped into three main classes, according to a veterinary classification: reproductive system diseases (i.e., metritis, retained foetal membranes, and ovarian cysts), locomotor system issues (i.e., lameness and generic leg infections), and gastroenteric diseases (i.e., abomasal displacement and dysentery). Other than the disease presence, no other information was available (e.g., no specific details on which type of mastitis or infection was diagnosed). In order to create a case-control dataset, for each disease, only the cows that manifested a disease at least once were kept in the dataset, hence removing all the animals that did not experience any disease during the study. Furthermore, all of the diseases were then summarised in a ‘generic disease’ variable, which described with 1/0 (i.e., presence/absence, respectively) the occurrence of at least one sanitary event.
Table 1. Recorded diseases and relative incidence in the data, in descending order
Statistical analysis
This study was composed of two main parts: in the first one, mixed models were used to analyse the effects of diseases on 2-h rumination time. All the models were fitted using the lme4 package (Bates et al. Reference Bates, Mächler, Bolker and Walker2015) in R (version 3.2.5; R Foundation for Statistical Computing, Vienna, Austria). Subsequently, the statistical significance of the model was checked with the lmerTest package (Kuznetsova et al. Reference Kuznetsova, Brockhoff and Christensen2015). Model 1 was fitted with the general disease variable:
$$\eqalign{{\rm Model (1)}\; \;rum\_mean_{ijk} & = generic\_disease_i \cr & + date_j + animal_k + \varepsilon _{ijk},} $$
where rum_mean ijk is the mean rumination time for animal k affected by a generic disease in test-day j; generic_disease i is the presence or absence of an unhealthy status; animal k is the random effect of the kth animal; date j is the random effect of the jth test day; and ε ijk is the random residual effect.
Model 2 was fitted including as independent variables each disease category:
$$\eqalign{{\rm Model}{\rm (}2)\; \;rum\_mean_{ijklmn} = & reprod_i + mast_j \cr & + locom_k + gastroent_l \cr & + animal_m + date_n + \varepsilon _{ijklmn},} $$
where rum_mean ijklmn is the mean rumination time for the animal m in the test-day n, affected or not by reprod i, mast j, locom k, and gastroent l; animal m and date n are the random effects; and ε ijklmn is the random residual effect.
In the second part of this study, a sliding windows approach was applied to the data to investigate the change in rumination time in a total of six different windows before and after the disease event (i.e., generic disease, reproductive diseases, mastitis, locomotor system issues, and gastroenteric diseases): the windows dimensions were of 1, 3, and 5 d, symmetrically set around the disease event. This approach is widely used in genomic analyses (e.g., linkage disequilibrium and signatures of selection identification), but is seldom applied outside of this field. On each window, the 2-h rumination mean, standard deviation (sd), and slope (from a linear regression of the rumination on the days in the window) were calculated. Furthermore, for each of these new parameters, summary statistics (i.e., mean ± sd) were calculated. Four different generalised linear models (Logistic regression) were then fitted on the window before the sanitary record, each with the disease event as a binary response (i.e., presence/absence: 1/0) and the afore-mentioned calculated parameters as predictors (Models 3.a to 3.d):
$$\eqalign{3{\rm (a)} disease &= rum\_mean \cr 3{\rm (b)} disease &= rum\_sd \cr 3{\rm (c)} disease &= rum\_slope \cr 3{\rm (d)} disease &= rum\_mean + rum\_sd + rum\_slope} $$
where disease is the presence or absence of one of the five cases analysed; rum_mean is the averaged rumination time in the window; rum_sd is the standard deviation of the rumination in the window; and rum_slope is the coefficient from the regression of the RT on the days in the window. AIC (Akaike information criterion) and McFadden's Pseudo R 2 (McFadden, Reference McFadden1974) were calculated to compare the models and assess which predictors and which window best fitted the data.
Results
Single variable comparison
The mean (±sd) RT of the animals in the herd, throughout the whole 377 d, was 46·99 ± 11·07 min/2 h. The effect of the disease presence on the 2-h RT was significant in every analysed case (P = 0·001 and P < 0·001, locomotor issues and all the other cases, respectively). Gastroenteric diseases had the largest effect, lowering RT by 9·91 min/2 h, while reproductive ones had the smallest, 1·08 min/2 h. Only three cows suffered from gastroenteric diseases, therefore, even if highly significant, the result must be interpreted with caution. The differences between the means (in min/2 h), the number of cows analysed, the ratio between positive and negative cases (case-control ratio), and the P-value from at-test, performed to assess if the differences between the two statuses were significant, are reported in Supplementary Materials, Table S1.
Multiple variables comparison
The fixed effects of Model 1 (estimated values, sem and P-values) are reported in Table 2, and their analysis of variance is reported in Table 3a. The random effects variances and standard deviations are reported in Table 3b. In this model, the diseased status had a significant effect (P < 0·001) on RT, lowering it by 2·22 min/2 h. The inclusion of the effect of the parity as predictor in the models was considered. However, with the inclusion of this effect in a preliminary test, the resulting model had the worst fit on our data (probably because this effect was confounded with the animal and date random effects), and parity was therefore removed.
Table 2. Models 1 and 2, 2-h interval
Fixed effects estimates, standard error of the means and significance by P-value
Table 3. Models 1 and 2, 2-h interval. (a) Statistical test for fixed effects. (b) Table of random effects, with their variances and standard deviations
The variance explained by the animal effect was 12·33 % of the total variance explained, and it was 6·30 times larger than the day effect variance. McFadden's Pseudo-R 2 of the model was 14·8 %.
The fixed effects of Model 2 (estimated values, sem and P-values are reported in Table 2) were statistically tested in the same way as Model 1, and the results are summarised in Table 3a, while the random effects variances and standard deviations are reported in Table 3b. The featured diseases in this model negatively affected RT, with a RT decrease which ranged from −1·73 to −5·76 min/2 h (reproductive and gastroenteric diseases, respectively). Similarly to the results of the general disease model, the variance explained by the animal effect (12·27 % of the total variance explained) was larger than the date effect (6·25 times larger). Pseudo-R 2 of the model was 14·8 %. Least Square Means (LSM) of 2-h RT by the different diseases are reported in Fig. 1. LSM were calculated for each single couple of diseased/non-diseased animals and were, therefore, different for different classifications. Variance inflation factor (VIF) was used as a diagnostic for multicollinearity within the models. In both of the models, multicollinearity was negligible. ANOVA was subsequently performed to assess which model (general disease variable vs. grouped diseases) better fitted the data: the two models were statistically different (P < 0·001).
Fig. 1. LSM of Rumination Time (RT) by specific disease presence (i.e., ‘diseased’) or absence (i.e., ‘non-diseased’), calculated for each single couple of diseased/non-diseased animals. The number of diseased animals for each disease is 126 (generic diseases), 66 (reproductive diseases), 68 (mastitis), 16 (locomotor issues), and 3 (gastroenteric diseases).
Sliding windows analysis
Regarding the sliding window analysis, summary statistics of the three disease predictors related to RT (i.e., mean, sd, and slope), for both the window before and after the disease event, are shown in Supplementary Materials, Table S2. In order to investigate if the pathological event changed the observed parameters ‘before’ and ‘after’ event occurrence, a t-test with a threshold of 0·05 for the P-value was used. For the generic disease analysis, only the slope was statistically different from before to after the event (different in all the three window's sizes). In the reproductive system diseases analysis, significant differences were identified only in the slope for windows’ sizes of 3 and 5 d. In the mastitis analysis, the 1 d window mean and all the slopes were statistically different. In the locomotor system issues analysis, the 5 d window mean and all the slopes excluding the one in 1 d window were statistically different. Lastly, in the gastroenteric diseases, a similar pattern as in the locomotor system issues were observed, i.e., all the slopes excluding the one in 1 d window were statistically different.
Logistic models were also fitted to the data: the estimate of the β, the odds ratio for the disease presence, the AIC of the model, and its Pseudo-R 2 are reported in Supplementary Materials, Table S3. In all of the five cases, the best model (i.e., lower AIC and higher Pseudo-R 2) was always Model 3.d, which fitted all the three considered predictors. For the generic disease analysis, mean, sd, and slope models showed significant effects in models from 3.a to 3.c (with a maximum Pseudo-R 2 of 2·99, 0·95, and 6·02 %, respectively), with the only exception of the 1 d window sd models, in which the effect is not significant. In Model 3.d, sd was never significant. Nevertheless, this model had the highest Pseudo-R 2 and the lowest AIC for all the three windows’ sizes. The reproductive system diseases model analysis showed a similar situation of the general disease analysis, although with lower Pseudo-R 2 values. Another important difference was the complete non-significance of all the models using sd as a predictor (Model 3.b). The mastitis model analysis had a similar pattern as the general disease one: the only non-significant window's size in the single-predictor models (i.e., Model 3.a to 3.c) was the Model 3.b, window's size of 1 d (sd). The maximum Pseudo-R 2 were 1·50, 1·14, and 4·10 %, respectively. Regarding Model 3.d, with window's size of one, sd was not significantly effective on RT. The locomotor system issues analysis showed a different pattern from the previous ones: the mean RT model (Model 3.a) had Pseudo-R 2 ten-fold higher than generic disease, reproductive diseases and mastitis ones. Similarly to the reproductive diseases analysis, however, Model 3.b was never statistically significant. Lastly, the gastroenteric diseases model analysis had, on average, the highest Pseudo-R 2 of all the analyses. The only non-significant window's size was the 1 d slope window. In the models with the three predictors (i.e., Model 3.d), both sd and slope were never statistically significant. Nevertheless, the Pseudo-R 2 ranged from 43·89 to 58·81 %.
The best models for each case, selected using the AIC and the Pseudo-R 2 as criteria, were always Model 3.d, but with a window's size of 5 d for the generic disease (Pseudo-R 2 = 6·47 %), a window's size of 3 d for reproductive diseases (Pseudo-R 2 = 7·16 %), a window's size of 1 d for mastitis (Pseudo-R 2 = 5·61 %), a window's size of 5 d for locomotor issues (Pseudo-R 2 = 16·17 %), and a window's size of 1 d for gastroenteric diseases (Pseudo-R 2 = 58·81 %).
Discussion
Automated rumination and activity monitoring could be used to identify diseases earlier than through clinical diagnosis performed by trained personnel. This confirms the results obtained by Stangaferro et al. (Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordano2016a, Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordanob, Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordanoc). All the comparisons between mean RT of diseased and non-diseased animals resulted statistically significant. Nonetheless, the differences were small and this type of comparison does not account for any interaction between variables and, therefore, it can identify large effects only. However, all of the different diseases’ effects were confirmed as statistically significant in the multiple variable approach too. The difference between the effects in the single and in the multiple variable comparison comes from the effect of the cow and the test-day, taken into account as random effects using the mixed model in the latter. These random effects should reduce the bias due to the correlation between the repeated measures.
The variance explained by the animal effect was larger than the date effect variance in both cases (i.e., general and grouped disease), suggesting that the observed variability is mainly due to the animal effect rather than to the test-day. This result is in accordance with the one from Byskov et al. (2015), where the authors observed that the 48 % of the total variation in RT was due to the animal effect, whereas feed intake accounted for the 32 %.
Compared through ANOVA, the two models were statistically different. Specifically, the model including specific variables for each disease fits the data better. Using mixed models, the effect of diseases on RT was confirmed in this study, though the model could be improved by adding further predictors (e.g., feed intake and diet of the animals, which was not available for this experiment). Results obtained with mixed models are in accordance with, and further expand, the results by Stangaferro et al. (Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordano2016a, Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordanob, Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordanoc).
With the sliding windows approach, we wanted to test if different features of RT in the days before a disease diagnosis could be predictive of the disease itself. This predictability would be desirable, since the detection of a disease as early as possible allows for a more immediate sanitary intervention. The features selected were the mean, the sd, and the regression slope of RT on time to disease. A difference in the slope before and after the diagnosis means that the rumination changes its trend (negative or positive). In the significant cases, the ‘after’ windows showed a positive (or less negative) trend, while the ‘before’ window had always a negative one: these results could suggest that these diseases affect rumination time, lowering it, and, since in our data the recorded event corresponds to the veterinary visit and the treatment beginning, we saw the improving of the rumination in the ‘after’ window due to medical treatment. From a descriptive point of view, sd of the RT was never statistically different before and after the disease event, while the mean showed a difference in the day before mastitis only. Of the selected feature, then, only the slope should be considered as a predictor, though the best model was always the one using all of the three features. The benefits from including mean and sd in the model are larger than the disadvantages, and this is probably due to a better representation of the phenomenon.
Different diseases could be predicted using different window size. Specifically, mastitis and gastroenteric diseases are better described by the models using one single day before the clinical diagnosis, while reproductive diseases and locomotor issues by the ones using 3 and 5 d, respectively. The Pseudo-R 2 of the reproductive diseases and mastitis predictive models was low. This could be due to the moderate ability to identify mild cases of metritis (Stangaferro et al. Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordano2016c) and mastitis caused by pathogens other than Escherichia coli through rumination changes. As stated by Stangaferro et al. (Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordano2016b), intramammary infections caused by E. coli are more easily identified because they are characterized by a severe inflammatory response, including sudden shock, sepsis, and often death. On the other hand, even with a small number of animals with gastroenteric disease, models predictive for these diseases had the highest Pseudo-R 2, in accordance with the high sensitivity detected by Stangaferro et al. (Reference Stangaferro, Wijma, Caixeta, Al-Abri and Giordano2016a). Locomotor issues, which were not analysed in the above-mentioned studies, showed a Pseudo-R 2 in between the other cases.
The difference in the window size in each different disease could be due not only to the higher or lesser effect of each disease on RT, but also to the different reaction time of the farmer in response to the different symptoms detected on his animals. Different diseases are perceived differently by farmers (e.g., mastitis is, from a commercial point of view, a greater concern than other diseases) and, therefore, they could require veterinary intervention with different urgency. Moreover, in order to obtain accurate estimates at enough distance from the event it is necessary for the farmer to carefully consider the occurrence of the events.
Conclusion
In this study we observed that common farm diseases (i.e., reproductive diseases, mastitis, locomotor system issues, and gastroenteric diseases) significantly affect the 2-h interval RT, lowering it in comparison to the one of healthy animals. Further studies are needed to fully assess the suitability of RT for predicting the onset of these diseases in individual animals. The growing presence of automatic recording systems, even in medium-small farms, will allow researchers to have larger datasets for modelling studies.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0022029917000619
The authors thank SCR Engineers Ltd. (Hadarim, Netanya, Israel) for supplying the technologies used in this study, Milkline Srl (Gariga di Podenzano, Piacenza, Italy) for the technical support and ‘Bulgarelli Giacomo e Astore’ dairy farm for the herd rumination raw data and health status recordings.
