Data for this study were collected from 10 dairy farms with 13 AMS. Three different AMS systems were used in study farms including Lely Astronaut (Lely, Rotterdam, Netherlands), DeLaval VMS (DeLaval, Tumba, Sweden) and Galaxy (Insentec, Marknesse, Netherlands). The mastitis detection systems of these AMS are based on sensors of electrical conductivity, milk yield and colour.

An initial characterization of the herds studied with respect to hygienic and productive variables was carried out. On all farms the cows were kept indoors and they had free access to a total mixed ration and in the AMS gained access to additional concentrate. Table 1 shows the number and types of AMS installed on each farm, the number of cows observed, linear score, days in milk, milk yield per cow and day, hygiene score for udders, thighs and legs and body condition score. Depending on the brand of AMS, the milking routines differed. Teats were cleaned with teat dipping cups or rotating brushes prior to milking. Milking cups were automatically attached immediately after and detached following milking and the teats were sprayed with a post-dipping bactericidal product. Teat cups were rinsed with warm water followed by water with or without peracetic acid or water vapour depending on the circumstances after each milking (Table 1).

Fourteen visits were made, one to 6 farms and two to the other 4 farms (Table 1). We visited the farms as close to the date of the monthly Official Milk Recording test (OFMRT) as possible (within 2 to 7 days). This way we could use the monthly test SCC data of each cow to validate other mastitis detection tests used in the study. At the beginning of each visit we collected the report of the milk quality alarms from the AMS mastitis detection software. These alarms, in general, provide alerts per cow and quarter of milk quality based on deviations of conductivity, colour, temperature (Hogeveen et al., 2010). With these two reports, we generated a list of all the animals with a possible infection around the time of the visit. In total 228 cows were analysed (912 milk quarters) and 176 were cows marked by AMS sensors with a milk quality problem. We used 52 cows with SCCs from all four quarters of approximately 1 million cells/mL as positive controls.

The CMT was performed on the 228 cows selected and was considered the gold-standard test for mastitis diagnosis. Samples were classified into 5 categories: when the mixture was visually normal (no gel-formation) it was scored as 0, no infection. If the reaction was weak, with traces that dissolved, the score was 1. In these cases we repeated the test to confirm the weak reaction. A weak thickening was scored as 2. A more severe thickening of the mixture but still able to spill by turning the paddle was a 3. And the formation of a gel such that the mixture stuck to the paddle was 4. All tests were performed by the same person. Postpartum cows were tested but the results were not included if less than 5 days had passed after calving because the results obtained from postpartum cows are difficult to interpret (Fouz et al., 2010).

Finally, milk samples were aseptically collected for bacteriological study from CMT-positive cows and were analysed at the Animal Health and Production Laboratory of Galicia, Lugo. Milk samples were cultured on Columbia agar plates containing 5% lamb blood, using a disposable sterile loop that seeded approximately 0.01 mL of milk using a laminar flow cabinet. Samples were incubated for 48 h at 37±2ºC and cultures were examined 24 and 48 h after incubation. Bacterial species were identified using biochemical API (Vitek 2, Biomerieux). A sample was considered contaminated if >2 bacterial species were isolated. When we suspected the presence of S. aureus the sample was subsequently cultured in selective BD Baird-Parker agar media. If after at least 72 h of incubation no microorganism was observed this sample was classified as having no bacterial growth.

The comparison of the milk quality alerts given by the AMS sensors with the CMT results was assessed using a classification model. If the CMT tested positive and was also classified by the AMS sensors as positive then it was considered a true positive (TP). When the AMS sensors and the CMT were negative the result was considered a true negative (TN). A false positive (FP) classification was a CMT negative quarter classified by the AMS sensors as positive. Finally, a false negative (FN) classification was a CMT positive quarter that was classified by the AMS sensors as negative. Using these four classifications, the detection of mastitis by AMS can be evaluated as follows: firstly the sensitivity as the fraction of CMT positive quarters classified as positive for mastitis by the AMS [Sensitivity (%) = 100 × TP / (TP + FN)]. And secondly, the specificity was defined as the fraction of CMT negative quarters classified as negative for mastitis by the AMS [Specificity (%) = 100 × TN / (TN + FP)]. The relationship between the benefits of AMS sensors (TP) and costs of a detection system (FP) was analyzed by a receiver operation characteristics (ROC) graph or Sensitivity vs. (1 – Specificity) plot (Kamphuis et al., 2010). The 14 predictions for the herd visits were plotted in the ROC graph. These allowed us to classify the herds based on their prediction methods.

Moreover, the success rate (SR) or predictive positive value was also calculated as SR = TP/(TP + FP). This represents the probability of the AMS correctly identifying a quarter as having mastitis. In addition we calculated the false alert rate as FAR = 1000 × FP/Total cows milked. Descriptive statistics were calculated for all variables.

Differences in sensitivity and prevalence between CMT score groups were contrasted by a one-way ANOVA with a Scheffe mean comparison. For categorical variables, distributions were analyzed using frequency tables. All data were processed using IBM SPSS 19.0.0 for Windows (SPSS, 2008).

Descriptive statistics for mastitis prevalence and the performance of AMS sensors used in the analysis are listed in Table 2. For 912 quarter milks used in the analysis 322 (35.3%) of them were infected quarters, determined by CMT test. The visual appearance of these 322 infected quarter milks on the CMT test, were different depending on the level of infection or SCC level. The frequency of abnormal milk with a CMT score of 1 was 40.1% with a variation throughout the 14 visits between 0 and 77.8% of infected quarter milk. A CMT score of 2 appeared in 36% of cases ranging from 13.6 to 90%. Variations within visits with a CMT score of 3 were also large with percentages ranging from 0 to 57.9% and a total frequency of 15.8%. The least common CMT score was 4 with a frequency of 7.8%. The prevalence rate of clinical mastitis varied from 0.04 to 0.14 per farm with a total of 0.09 (Table 2).

The AMS classified as positive mastitis cases 210/912 (23.0%) quarters but 36 of these were CMT negative and considered false positives (Table 2). Among the remaining 702 AMS negative quarters, 16.2% were CMT positive and therefore false negatives. So average sensitivity of the mastitis detection systems of the AMS studied was 58.2% and the specificity was 94%. The positive prediction value (PPV) ranged from 0.53 to 1 with a mean of 0.86, while the negative prediction value (NPV) was of 0.80. The average FAR was 37.8% of the cases (Table 2). The specificity in all cases is greater than 80% but only on three of the visits was the sensitivity above 70%. Most false negative cases (52.7%) are associated with a CMT score of 1 (Fig. 1) however, differences in sensitivity in the CMT 1, 2, 3 and 4 categories were not statistically significant (p = 0.104) (Table 3). However, the prevalence values were significantly different between categories of CMT (p = 0.008).

The total samples analyzed were 181. The percentage of quarters with specific mastitis pathogens identified was 31.5% (57/181) Pathogens were classified as environmental (18%) and contagious (13.8%) of which 10.5% were secondary contagious pathogens (Table 4). However, the most prevalent pathogen isolated, Streptococcus dysgalactiae (28.1%), was contagious. Streptococcus uberis was found in 26.3% of samples with isolated pathogens (8.3% of all samples) followed by S. aureus (10.5%). No bacteria were isolated in 30.9% of samples and 18.8 % of them were considered to be contaminated (Table 4).

There are few studies estimating the prevalence of mastitis at milk quarters level instead of cow level. The prevalence of quarters with mastitis on the present work with AMS was 9%, lower than that in Dimitar & Metodija (2012) study, where prevalence was also analysed at milk quarter level (15%). Although in terms of mastitis detection, the minimum recommendations are a sensitivity of 80% and a specificity of 99% (Hogeveen et al., 2010), our AMS mastitis detection system did not reach these figures and our results do not agree with other researchers who reported higher values for specificity and sensitivity in both conventional milking systems (using single quarter samples) (Lam et al., 2009) and in AMS (Steeneveld et al., 2010a). However, these minimum levels for specificity and sensitivity are still under discussion (Hogeveen et al., 2010). Results suggest that the performance of mastitis detection systems is similar to other regions with the same milking system (Steeneveld et al., 2010b) where these authors claim that a sensitivity of 43% and an specificity of 97% are the clinical mastitis detection characteristics of current AMSs, like it was found from visual inspections of all milk samples obtained during three days on three Dutch commercial dairy farms (Mollenhorst & Hogeveen, 2008). Some authors consider that it is impossible for AMS systems to have a sensitivity of 100% (Kamphuis et al., 2010). An increase in the SR and a decrease in the FAR were confirmed when in-line SCC information was added to a detection model of AMS using electrical conductivity information (Kamphuis et al., 2008). According to our study half of the false negative cases were in milk quarters classified with the lowest CMT. This could mean that either lowest levels of infection are difficult to detect by the sensors installed in the AMS, or that there may not have been a real infection and that the error was in our interpretation of the CMT test. CMT interpretation can be subjective (Polat et al., 2010), and this might result in false positives and negatives (Viguier et al., 2009) when not being executed correctly (Lam et al., 2009). Some authors argue that quarters with low CMT-scores in foremilk probably do not have clinical mastitis (Rasmussen, 2001) or they could be subclinical cases. However, more severe mastitis cases are less frequent but easier to detect by the AMS sensors. If the reaction was weak, with “traces” that dissolved (scored as 1) even we had doubts about a possible mastitis because they could be cows with high days in milk (DIM) with a high cell count which can react with the CMT due to cell flaking (Fouz et al., 2004).

The contaminated samples together with the negative results and the ones in which no microorganisms were isolated represent more than half of the samples. This high percentage highlights the importance of a proper procedure when taking samples. These data coincide with that shown in a study, which also took place in Galicia but, in dairies with conventional milking systems (Cundins et al., 2010). However, the percentage of contaminated samples was higher than that of Olde Riekerink et al. (2008) study. Similar data were shown, with respect to the lack of bacterial growth in CMT-positive milk, in Makovek & Ruegg (2003) and Oliveira et al. (2013). The high rate of samples without results or contaminated samples was caused by the difficulty in sampling, due to the situation in which it is performed. To effectively use bacteriological culturing as a diagnostic tool, milk samples have to be collected from the correct cows and quarters at the correct point in time (Lam et al., 2009). In other cases it is possible that the time needed for bacterial growth was higher than the time allowed in the present study. It can also occur that the culture media used was not suitable for a specific pathogen as shown for S. aureus (Fouz et al., 2004).

S. aureus can be the pathogen with the greatest incidence in mastitis cases (Olde Riekerink et al., 2008), nonetheless, in our study it was found in only 3.3% of total samples. In many other studies the most common mastitis-causing agents were the coagulase-negative staphylococci (Ferguson et al., 2007; Nam et al., 2010; Schwarz et al., 2010). Although environmental mastitis pathogens were the most common causative agents, it was S. dysgalactiae which was the most particular pathogen in this study, being a contagious pathogen. Transfer of bacteria by the AMS pre-milking teat-cleaning device was suspected to be one cause for increased infections (Hovinen & Pyörälä, 2011). However, in this study the level of hygiene may be considered adequate since the percentage of pathogens related to udder hygiene such as Klebsiella spp. and E. coli was very low. In fact one Finnish study using AMS had a greater prevalence of these pathogens compared to the present study (Hovinen et al., 2005). Moreover S. uberis had a relatively high prevalence in our study and these bacteria can be considered as being both a contagious and an environmental agent.

In conclusion, dairy farms with AMS in this study had a similar prevalence of mastitis and pathogen profile as farms with conventional milking systems. The majority of the bacteria isolated from these herds were environmental pathogens and special attention needs to be placed on the prevention and control of environmental and contagious mastitis pathogens as all of the cows are milked with the same machine and with AMS, milk cups are not disinfected between cows. Results suggest that the performance of an AMS mastitis detection system was similar to that in other regions with the same milking system, and lower than the CMT. More severe mastitis cases are less common but easier to detect with AMS mastitis detection systems.