Research trends and evidence landscape of metabolomics and machine learning in the diagnosis of subclinical ketosis in dairy cows: a systematic review and bibliometric analysis.

Subclinical ketosis (SCK) is a common metabolic disorder, affecting 20% to 40% of postpartum dairy cows depending on diagnostic matrix (blood vs. milk ketone testing) and days in milk (DIM), compromising milk yield and reproduction. Conventional diagnostics rely on blood β-hydroxybutyrate (BHB) thresholds lacks predictive capability for early detection. Metabolomics and machine learning (ML) provides promising non-invasive approaches by analyzing biofluid metabolites and complex datasets. This systematic review and bibliometric analysis evaluated global research trends and the evidence landscape of metabolomics and ML applications in SCK diagnosis. Following a PRISMA 2020-adapted search of Scopus (2013-January 2026), 160 records were identified, of which 21 peer-reviewed studies met the inclusion criteria. Methodological quality, evaluated using QUADAS-2, showed mixed risk of bias, particularly in animal selection, index test, reference standards, flow and timing domains. Bibliometric analysis examined the annual production, most global cited documents, core journals, local impact of core journals, three-field plot, keyword co-occurrence, trend topics, thematic clustering, and evolution. Publications accelerated after 2020, with European dominance and milk infrared spectroscopy as the primary metabolomics platform (n = 11). Machine learning methods included random forest (n = 8) and artificial neural networks (n = 5), achieving accuracies of 68.0%-85.7% (AUC = 0.77-0.89) for SCK prediction, offering practical utility as a non-invasive screening tool for herd-level risk monitoring. Key themes centered on lactation stage, biomarkers, and non-invasive monitoring, with the emerging integration of multi-omics and sensor data. Overall, metabolomics and ML provide preliminary screening insights for herd-level monitoring, supporting conventional diagnostic protocols. However, persistent limitations, regional bias, small/single-herd datasets, heterogeneous thresholds, and limited multi-omics/longitudinal designs constrain generalizability and translational impact. Future research should prioritize multi-center validation, standardized protocols, and explainable ML models to support practical on-farm implementation.