When the hive speaks: detecting Varroa destructor infestation in honey bees through acoustic biomarkers and machine learning.

Honey bee (Apis mellifera) colonies are essential for agricultural production and ecosystem services but are increasingly threatened by multiple stressors, particularly the parasitic mite Varroa destructor, one of the major causes of global colony losses. Early and reliable detection of infestations is therefore critical for effective colony management. In this study, we investigated continuous hive acoustics as a non-invasive approach for the early detection of Varroa infestations. Approximately 960,000 timestamped acoustic recordings were collected from 53 colonies, of which only recordings temporally matched to available Varroa assessments (May-August 2020) were retained for model development. Each recording was temporally aligned with field-based Varroa measurements, and classification models were developed using 19 acoustic features derived from frequency bands and signal intensity. Among the tested algorithms, Random Forest achieved the highest performance (accuracy: 89.7%, AUC: 0.964), followed by XGBoost (82.4%, 0.908) and LightGBM (77.0%, 0.866). Permutation importance analysis revealed that low- to mid-frequency components, particularly within the 122-396 Hz range, served as sensitive acoustic indicators of Varroa presence. These findings demonstrate that hive acoustics capture biologically meaningful acoustic signatures associated with Varroa infestation status and highlight their potential as a scalable, non-invasive decision-support tool for early detection and improved colony management in apiculture.