DOI: 10.56430/japro.1979815 ISSN: 2757-6620

Artificial Intelligence-Assisted Disease and Welfare Detection in Farm Animals: A Diagnostic Test Accuracy Systematic Review and Meta-Analysis

Muhammet Kuddusi Erhan, Aycan Mutlu Yağanoğlu
Show PDF Cite
To estimate, against a reference standard, how accurately artificial intelligence methods detect disease and welfare conditions in farm animals. PubMed, Web of Science, and Scopus were searched for primary diagnostic accuracy studies, with an updated search during revision for additional animal-level lameness studies. When a 2×2 table was not reported, the table was reconstructed from the reported metrics and checked for internal consistency. Eligible studies were pooled with a bivariate random-effects model, and the results were checked with independent estimators. Heterogeneity, publication bias, sensitivity analyses, risk of bias, and certainty of evidence were assessed. Seventy-four of 154 full texts met the inclusion criteria. Under an eligibility criterion amended in review to keep within-group sample sizes comparable (animal-level N ≤ 500 cows), three large-sample studies were excluded and one small-sample study was added through the updated search; ten studies entered the primary meta-analysis (seven on mastitis and three on lameness). A single-study reproduction/metritis dataset was excluded from pooling and synthesised narratively. Pooled sensitivity was 83.5% (95% confidence interval 74.0–90.0) and pooled specificity was 90.1% (95% confidence interval 85.4–93.3), with an area under the curve of 0.936 and a diagnostic odds ratio of 46. The mastitis subgroup (k=7) reached 83.5% sensitivity and 88.9% specificity; the three-study lameness subgroup is reported as exploratory. Heterogeneity was substantial, and most studies could not be pooled because they reported performance at the record or quarter level rather than per animal. Artificial intelligence-based detection appears promising, combining high specificity with good sensitivity, but wide heterogeneity, the small number of poolable studies, and low certainty of evidence all call for cautious interpretation.

More from our Archive