Akademik Veri Yönetim Sistemi
Kafkas Universitesi Veteriner Fakultesi Dergisi, cilt.32, sa.1, ss.141-146, 2026 (SCI-Expanded, Scopus, TRDizin)
Pollinator populations, which play a critical role in maintaining global ecosystem health, have been experiencing marked declines worldwide due to widespread pesticide usage. However, early behavioral indicators of lethal stress induced by chemical exposure remain insufficiently characterized, largely because conventional ecotoxicological assessments predominantly focus on mortality-based endpoints. In this study, we evaluated the potential to predict mortality risk at an early stage using behavioral markers, based on 1.506 behavioral observation records collected from seven bee species exposed to lambda-cyhalothrin. To this end, we implemented explainable artificial intelligence models, including Random Forest, XGBoost, and LightGBM, and interpreted the model outputs using SHAP analysis. Among these models, Random Forest and XGBoost demonstrated the strongest performance in distinguishing high mortality risk, achieving an accuracy of 0.873 on an independent test dataset. SHAP-based model interpretation revealed a temporal behavioral progression associated with elevated mortality risk: cramps and apathy emerged as early warning indicators (2–4-hour window), uncoordinated movement represented the intermediate phase, and the dorsal recumbent position characterized the terminal collapse stage. These findings demonstrate that behavioral early-warning signals of lethal pesticide stress can be reliably detected prior to mortality and highlight the potential of explainable artificial intelligence as a robust decision-support tool for pollinator health monitoring and pesticide risk assessment.