Prioritizing alarms from sensor-based detection models in livestock production - A review on model performance and alarm reducing methods

Sensor-based detection systems for decision support generate too many false alarms.Methods for reducing or prioritizing alarms are described and evaluated.Performance criteria based on sensitivity and specificity are not fulfilled.Three included papers present prioritizing strategies for reducing alarms.Alternative approaches for communicating alarms in future research are proposed. The objective of this review is to present, evaluate and discuss methods for reducing false alarms in sensor-based detection models developed for livestock production as described in the scientific literature. Papers included in this review are all peer-reviewed and present sensor-based detection models developed for modern livestock production with the purpose of optimizing animal health or managerial routines. The papers must present a performance for the model, but no criteria were specified for animal species or the condition sought to be detected. 34 papers published during the last 20years (19952015) are presented in three groups according to their level of prioritization: Sheer detection models based on single-standing methods with or without inclusion of non-sensor-based information (19 papers), Improved detection models where the performance of the described models are sought to be improved through the combination of different methods (12 papers) and Prioritizing models where the models include a method of ranking or prioritizing alerts in order to reduce the number of false alarms (3 papers). Of the three methods that rank or prioritize alerts; Fuzzy Logic, Naive Bayesian Network (NBN) and Hidden phase-type Markov model, the NBN shows the greatest potential for future reduction of alerts from sensor-based detection models in livestock production. The included detection models are evaluated on three criteria; performance, time-window and similarity to determine whether they are suitable for implementation in modern livestock production herds. No model fulfills all three criteria and only three models meet the performance criterion. Reasons for this could be that both sensor technology and methods for developing the detection models have evolved over time. However, model performance is almost exclusively presented by the binary epidemiological terms Sensitivity (Se) and Specificity (Sp). It is suggested that future research focus on alternative approaches for the output of detection models, such as the prior probability or the risk of a condition occurring. Automatic monitoring and early warning systems offer an opportunity to observe certain aspects of animal health, welfare, and productivity more closely than traditionally accomplished through human observation, and the opportunities for improving animal welfare should continue to be a driving force throughout the field of precision livestock farming.