Making sense of sensor data: detecting clinical mastitis in automatic milking systems

Farmers milking dairy cows are obliged to exclude milk with abnormal homogeneity or color for human consumption (e.g., Regulation (EC) No 853/2004), where most abnormal milk is caused by clinical mastitis (CM). With automatic milking (AM), farmers are no longer physically present during the milking process. AM systems use sensors in order to replace the visual monitoring of udder health. Sensor measurements are used by CM detection models as input data to produce mastitis alert lists. These lists report cows suspected of having CM as they deviate, according to the sensor information, from ‘normal’ for one reason or another. It is the responsibility of the dairy farmer to check listed cows visually to confirm CM. Most CM detection models use sensor information of the electrical conductivity as input data. Currently used CM detection models, however, can be improved in order to detect more CM cases and to decrease the number of cows listed on the mastitis list erroneously. The objective was to improve CM detection using sensor data from AM systems by exploring three routes: (1) applying new algorithms, (2) adding new sensor information, and (3) adding non-sensor information. This improved CM detection model should (1) detect CM at high levels of sensitivity (>70%) and specificity (>99%), (2) alert for CM preferably at the quarter milking where clinical signs are visible for the first time, or within a limited time period around this event, (3) be able to deal with field data, and (4) provide information about the CM causal pathogen in order to improve the decision making process regarding antibiotic treatment. Nine Dutch dairy farmers participated in a field study to collect sensor data and visual observations of CM. Sensor measurements were transformed to predictive variables, which were used as input data for the development of a CM detection model with decision tree induction. Detection performance of the developed model (sensitivity 40% at a specificity 99%) was higher compared with models currently used by AM systems. It got close to the required performance levels by increasing the time window in which the model was allowed to alert for a CM event (sensitivity 69.5%, specificity 99%). Sensor data also had potential to differentiate between Gram-positive and Gram-negative CM causal pathogen. Adding information from a new sensor (measuring somatic cell count (SCC)) to a CM detection model that uses sensor measurements of the electrical conductivity improved the detection of CM as well. Despite expectations, adding non-AM cow information (e.g., parity, SCC history, CM history) did not improve detection performance when added to a detection model based on sensor data. The next step in improving CM detection models is gaining insight in the needs of dairy farmers, obtaining international agreements on gold standard definitions, applied time windows, and data inclusion criteria, and collecting a data set from more than one commercial dairy farm where all quarter milkings are checked visually for CM.