C. Kamphuis

Sensors and Clinical Mastitis—The Quest for the Perfect Alert

When cows on dairy farms are milked with an automatic milking system or in high capacity milking parlors, clinical mastitis (CM) cannot be adequately detected without sensors. The objective of this paper is to describe the performance demands of sensor systems to detect CM and evaluats the current performance of these sensor systems. Several detection models based on different sensors were studied in the past. When evaluating these models, three factors are important: performance (in terms of sensitivity and specificity), the time window and the similarity of the study data with real farm data. A CM detection system should offer at least a sensitivity of 80% and a specificity of 99%. The time window should not be longer than 48 hours and study circumstances should be as similar to practical farm circumstances as possible. The study design should comprise more than one farm for data collection. Since 1992, 16 peer-reviewed papers have been published with a description and evaluation of CM detection models. There is a large variation in the use of sensors and algorithms. All this makes these results not very comparable. There is a also large difference in performance between the detection models and also a large variation in time windows used and little similarity between study data. Therefore, it is difficult to compare the overall performance of the different CM detection models. The sensitivity and specificity found in the different studies could, for a large part, be explained in differences in the used time window. None of the described studies satisfied the demands for CM detection models.

Automatic detection of clinical mastitis is improved by in-line monitoring of somatic cell count

This study explored the potential value of in-line composite somatic cell count (ISCC) sensing as a sole criterion or in combination with quarter-based electrical conductivity (EC) of milk, for automatic detection of clinical mastitis (CM) during automatic milking. Data generated from a New Zealand research herd of about 200 cows milked by 2 automatic milking systems during the 2006-2007 milking season included EC, ISCC, monthly laboratory-determined SCC, and observed cases of CM that were treated with antibiotics. Milk samples for ISCC and laboratory-determined SCC were taken sequentially at the end of a cow milking. Both samples were derived from a composite cow milking obtained from the bottom of the milk receiver. Different time windows were defined in which true-positive, false-negative, and false-positive alerts were determined. Quarters suspected of having CM were visually checked and, if CM was confirmed, sampled for bacteriological culturing and treated with an antibiotic treatment. These treated quarters were considered as gold-standard positives for comparing CM detection models. Alert thresholds were adjusted to achieve a sensitivity of 80% in 3 detection models: using ISCC alone, EC alone, or a combination of these. The success rate (also known as the positive predictive value) and the false alert rate (number of false-positive alerts per 1,000 cow milkings) were used to evaluate detection performance. Normalized ISCC estimates were highly correlated with normalized laboratory-determined SCC measurements (r = 0.82) for SCC measurements >200 x 10(3) cells/mL. Using EC alone as a detection tool resulted in a range of 6.9 to 11.0% for success rate, and a range of 4.7 to 7.8 for the false alert rate. Values for the ISCC model were better than the model using EC alone with 12.7 to 15.6% for the success rate and 2.9 to 3.7 for the false alert rate. Combining sensor information to detect CM, by using a fuzzy logic algorithm, produced a 2- to 3-fold increase in the success rate (range 21.9 to 32.0%) and a 2- to 3-fold decrease in the false alert rate (range 1.2 to 2.1) compared with the models using ISCC or EC alone. Results suggest that the performance of a CM detection system improved when ISCC information was added to a detection model using EC information.

Detection of clinical mastitis with sensor data from automatic milking systems is improved by using decision-tree induction

The objective was to develop and validate a clinical mastitis (CM) detection model by means of decision-tree induction. For farmers milking with an automatic milking system (AMS), it is desirable that the detection model has a high level of sensitivity (Se), especially for more severe cases of CM, at a very high specificity (Sp). In addition, an alert for CM should be generated preferably at the quarter milking (QM) at which the CM infection is visible for the first time. Data were collected from 9 Dutch dairy herds milking automatically during a 2.5-yr period. Data included sensor data (electrical conductivity, color, and yield) at the QM level and visual observations of quarters with CM recorded by the farmers. Visual observations of quarters with CM were combined with sensor data of the most recent automatic milking recorded for that same quarter, within a 24-h time window before the visual assessment time. Sensor data of 3.5 million QM were collected, of which 348 QM were combined with a CM observation. Data were divided into a training set, including two-thirds of all data, and a test set. Cows in the training set were not included in the test set and vice versa. A decision-tree model was trained using only clear examples of healthy (n=24,717) or diseased (n=243) QM. The model was tested on 105 QM with CM and a random sample of 50,000 QM without CM. While keeping the Se at a level comparable to that of models currently used by AMS, the decision-tree model was able to decrease the number of false-positive alerts by more than 50%. At an Sp of 99%, 40% of the CM cases were detected. Sixty-four percent of the severe CM cases were detected and only 12.5% of the CM that were scored as watery milk. The Se increased considerably from 40% to 66.7% when the time window increased from less than 24h before the CM observation, to a time window from 24h before to 24h after the CM observation. Even at very wide time windows, however, it was impossible to reach an Se of 100%. This indicates the inability to detect all CM cases based on sensor data alone. Sensitivity levels varied largely when the decision tree was validated per herd. This trend was confirmed when decision trees were trained using data from 8 herds and tested on data from the ninth herd. This indicates that when using the decision tree as a generic CM detection model in practice, some herds will continue having difficulties in detecting CM using mastitis alert lists, whereas others will perform well.

Sensor data on cow activity, rumination, and ear temperature improve prediction of the start of calving in dairy cows

Adding sensor data to a predictive model improves prediction of the calving day.Sensor detected change in activity had most value to predict the start of calving.The predictive model predicts 43.5% of the calvings with 1% false positive alerts.Notable change in sensor data was observed within 10h before calving started.The predictive model generated most alerts in the last 12h prior to calving. Management during calving is important for the health and survival of dairy cows and their calves. Although the expected calving date is known, this information is imprecise and farmers still have to check a cow regularly to identify when it starts calving. A sensor system that predicts the moment of calving could help farmers efficiently check cows for calving. Observation of a cow prior to calving is important because dystocia can occur, which requires timely intervention to mitigate adverse effects on both cow and calf. In this study, 400 cows on a Dutch dairy farm were equipped with sensors. The sensor was a single device in an ear tag, which synthesised cumulative activity, rumination activity, feeding activity, and temperature on an hourly basis. Data were collected during a one-year period. During this period, the starting moment of 417 calvings was recorded using camera images of the calving pen taken every 5min. In total, 114 calving moments could be linked with sensor data. The moment at which calving started was defined as the first camera snapshot with visible evidence that the cow was having contractions or had started labor. Two logit models were developed: a model with the expected calving date as independent variable and a model with additional independent variables based on sensor data. The areas under the curves of the Receiver Operating Characteristic were 0.885 and 0.929 for these models, respectively. The model with expected calving date only had a sensitivity of 9.1%, whereas the model with additional sensor data has a sensitivity of 36.4%, both with a fixed false positive rate of 1%. Results indicate that the inclusion of sensor data improves the prediction of the start of calving; therefore the sensor data has value for the prediction of the moment of calving. The model with the expected calving date and sensor data had a sensitivity of 21.2% at a one-hour time window and 42.4% at a three-hour time window, both with a false positive rate of 1%. This indicates that prediction of the specific hour in which calving started was not possible with a high accuracy. The inclusion of sensor data improves the accuracy of a prediction of the start of calving, compared to a prediction based only on the expected calving date. Farmers can use the alerts of the predictive model as an indication that cows should be supervised more closely in the next hours.