Machine learning for detecting colony collapse disorder and colony losses in honeybees (Apis mellifera): a PRISMA 2020 systematic review

Abstract Introduction: Globally, pollination services and beekeeping livelihoods are at risk due to colony collapse disorder (CCD) and increased colony losses. Although early warning is promised by machine learning (ML) applied to hive sensor, image, and survey data, the proof has not been synthesized with a concentration on collapse and loss. Objectives: This systematic review mapped and assessed studies that employed machine learning (ML) to identify or predict CCD, colony loss, or the colony-health states that precede loss in Apis mellifera, commonly known as honeybees. The data, models, validation processes, and deployment levels were also examined. Methods: Peer-reviewed studies published between 2015 and 2025 were found using citation searches of two recent reviews and organized academic web searching after PRISMA 2020. 58 studies were included after 110 records were screened and 69 full texts were evaluated. An eight-point standardized checklist tailored for machine learning research was used to evaluate methodological quality after task, modality, dataset, model, validation, performance, and deployment data were extracted. Results: Colony collapse was explicitly modeled in only two studies. The remaining investigations identified the following precursors: swarming (8), multi-condition anomalies (12), parasites and diseases (16 studies, primarily Varroa destructor), queen status (13), colony loss or winter mortality (4), and colony strength (3). Deep learning was the most common model family (20 research), and acoustic monitoring predominated (27 studies). Although just ten research examined models on independent hives, where performance significantly declined, reported accuracies frequently approached 95%; prediction of loss models reached roughly 65–75% accuracy. 18 studies were good quality, 26 had moderate quality, and 14 had low quality. Conclusions: Although there is little evidence to forecast collapse itself, machine learning (ML) consistently identifies a few proximate causes of colony failure. Multimodal fusion, hive-independent validation, open data and code, and longitudinal, multi-apiary datasets with confirmed loss consequences are all necessary for progress.