Comparative performance analysis of three machine learning algorithms applied to sensor data registered by a leg-attached accelerometer to predict metritis events in dairy cattle

Routinely collected sensor data could be used in metritis predictive modeling but a better understanding of its potential is needed. Our objectives were 1) to compare the performance k -nearest neighbors ( -NN), random forest (RF), and support vector machine (SVM) classifiers on detection behavioral patterns associated with events measured by leg-attached accelerometer (TrackaCow, ENGS, Hampshire, UK); 2) study impact farm scheduled activities model performance; 3) identify which behaviors yield highest F 1 score for prediction as function number time window time-lags. A total 239 (188 non-metritis 51 events) retrospectively created based changes two consecutive uterine evaluations from dataset containing clinical during first 21 days postpartum between June 2014 May 2017. These 10,874 - 14,138 points corresponding hourly measurements lying time, bouts, steps, intake, intake visits. Sensor 3 before each event aggregated every 24-, 12-, 6-, 3-hour windows. Multiple time-lags also determine optimal past observations needed classification. Similarly, different decision thresholds compared. Depending classifier, algorithm hyperparameters optimized using grid search (RF, -NN, SVM) (RF). All changed throughout period showed distinct daily patterns. From three algorithms, RF had score, no classifier performance. Furthermore, 3- 6-hour windows best balance scores We concluded that steps can predict 2 event. Findings this will develop more complex models cows at higher risk experiencing metritis.