Machine learning for detecting colony collapse disorder and colony losses in honeybees (Apis mellifera): a PRISMA 2020 systematic review
This PRISMA 2020 systematic review of 58 studies reveals that while machine learning effectively identifies proximate causes of honeybee colony failure like parasites and queen status using acoustic and deep learning methods, it currently lacks robust evidence for directly predicting colony collapse disorder, necessitating future research focused on independent validation, multimodal data fusion, and open longitudinal datasets.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Honeybees are the world's most vital pollinators, moving pollen between crops and wild plants to ensure the growth of food and flowers. For decades, beekeepers have watched with growing alarm as entire colonies vanish or dwindle, a phenomenon once called colony collapse disorder and now understood as a complex mix of stressors including parasites, pesticides, and climate extremes. Because checking a hive by hand is difficult, disruptive, and often too late to save the bees, scientists have turned to technology. They attach sensors to hives to listen to the bees, weigh the boxes, and watch for changes in temperature or sound. The hope is that computers can learn to recognize the subtle signs of a dying colony long before the bees disappear, acting as an early warning system.
A recent systematic review by Fredrick Michael Ogore at Daystar University in Nairobi brings together fifty-eight studies from around the world to see how well this promise is being kept. The researcher gathered every peer-reviewed paper published between 2015 and 2025 that used machine learning to detect honeybee colony losses. The goal was not just to count the studies, but to understand what they actually found, how they tested their ideas, and whether the technology is ready to help real beekeepers. The review reveals a field that is moving fast, with a heavy shift toward deep learning and edge computing, yet it also highlights a significant gap between the high scores reported in laboratories and the reality of predicting a colony's death.
The most striking finding is that very few studies actually tried to predict the total collapse of a colony. Instead, the vast majority of research focused on spotting the immediate causes that lead to failure. The most common targets were parasites, specifically the Varroa mite, which was the subject of sixteen studies. Researchers used cameras and computer vision to count mites on bees or on sticky boards, with some systems achieving high accuracy in controlled settings. Another thirteen studies focused on detecting whether a queen bee was present, while eight looked for the specific sounds of swarming. These are all critical problems, but they are precursors rather than the final event of a colony dying. Only two studies explicitly modeled the collapse itself, and even those relied on limited data.
The data used to train these computer models comes mostly from Europe, with significant contributions from the Czech Republic, the United Kingdom, and Spain. The most popular method for gathering information is listening to the hive. Twenty-seven studies used audio recordings, analyzing the hum and buzz of the bees to identify problems like a missing queen or an impending swarm. In many of these acoustic studies, the computers reported accuracy rates approaching 95% or even 99%. However, the review suggests these numbers are often misleading. Most studies tested their models on data from the same few hives used to train them. This is like testing a student only on the exact questions they practiced; it does not prove they can answer new questions. When researchers tested their models on entirely different hives, the performance dropped significantly, often falling to around 65% to 75% for predicting actual colony loss.
The quality of the research varied widely. The reviewer rated the studies on an eight-point checklist, finding that only eighteen were of high quality. Many studies lacked crucial details, such as testing on independent hives or sharing their data and code for others to verify. While some researchers have successfully built small, low-power devices that can listen to bees and make decisions right on the hive without needing the internet, these systems have not yet been tested in long-term field trials with real beekeepers. There is no evidence yet that these alarms reduce the number of dead colonies or that they work across different climates and bee subspecies.
Ultimately, the review concludes that while machine learning is excellent at identifying the immediate symptoms of a struggling hive, such as a missing queen or a heavy mite infestation, it is not yet a reliable crystal ball for predicting total colony failure. The technology has moved from theory to practical tools like smartphone apps for counting mites and edge devices for listening to bees, but the leap to a robust early-warning system for collapse remains unfinished. To get there, the scientific community needs to move beyond single-hive experiments and build large, shared datasets that track colonies over time with confirmed outcomes. Only then can these digital tools be trusted to save the bees before it is too late.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.