Review: Establishing precision, bias, and reproducibility standards for dairy cattle behavior sensors

Abstract

This scoping review addressed the disparity in statistical approaches for validating wearable sensors in dairy cattle behavior research. The objective of this scoping review was to (1) synthesize 101 original research articles that validated wearable sensors to observe dairy cattle behavior (activity and feeding behavior) from the past 11 years to build a reference point for researchers, (2) make recommendations for statistical reporting (precision, bias, and minimum reporting standards) for future validation research that uses wearable sensors to record dairy cattle behavior, focusing on calculating precision, and bias, and reporting reproducibility criteria, and (3) evaluate which validated wearables met our criteria for validity; ≥ 85% precision, reported repeatability criteria, and no bias was observed. A systematic search across PubMed and Web of Science yielded 2 955 articles, which were reduced to 101 after duplicate removal. Data extraction, performed with Power BI, classified accuracy, precision, bias, sensitivity, specificity, reproducibility, sensor type, gold standard, and observed behaviors. Precision was defined as a calculated precision value or the use of Lin's Concordance Correlation Coefficient (CCC), Pearson correlations, or linear regressions. A study was considered precise if it demonstrated > 85% precision or high correlations/explained variability (≥ CCC or R² = 0.85). Bias was identified through Bland-Altman plots, deviations from the mean, best-fit lines, bias correction factors, or location scale shifts. Reproducibility required defining sensor type, sample size, commercial name, and behaviors in an ethogram table. Validity criteria mandated that a study be precise, reproducible, and exhibit no bias. Activity behavior was the most frequently studied (61/101), followed by consumption (59/101), resting (55/101), and digestive behaviors (49/101). A high proportion, 93% (94/101), met reproducibility criteria. However, only 40% (40/101) calculated precision or used a proxy. Of those reporting precision, 90% (36/40) were precise, and 95% (38/40) were reproducible, but only 35% (14/40) reported bias. Overall, only 14% (14/101) of the reviewed studies met all validity criteria. Validated behaviors included activity, feeding, and consumption. Sensors meeting validity criteria were IceCube, Nedap Smart Tag, RumiWatch, Smartbow GmbH, MooMonitor+, Hobo Pendant G, and CowManager. Future validation studies should prioritize calculating precision, reporting sample size and sensor details, and statistically assessing bias to ensure reliable data for dairy farmers.

Keywords: Animal welfare; Automated behavior monitoring; Precision livestock farming; Sensor validation metrics; Smart farming.