Automated pose estimation and machine learning for classifying insecticide modes of action

Insect behavior is governed by intricate neuromuscular systems, thereby abnormal behaviors can be induced by neurotoxic insecticides. Traditional approaches such as RFID and color markers are limited to tracking movement trajectories, which hardly distinguish among different insecticide classes. This study employs automated pose estimation guided by Convolutional Neural Networks (CNN) to track the simultaneous position of multiple body parts in the recorded video. By analyzing the relative geometric configuration of these body parts, we obtained detailed information on the alterations in motor activity following insecticide exposure. We treated insects with several representative classes of insecticides to derive altered motor activity parameters. Deep learning-based classification algorithms, including Random Forest, eXtreme Gradient Boosting, and K-Nearest Neighbors, were then used to determine if each mode of action has distinct behavioral abnormality patterns. Additionally, we tested whether such detailed symptomatic analysis can infer the modes of action of randomly provided compounds. In an experiment where three insecticides were randomly provided to the experimenter, the mode of action was speculated solely using the described framework. The results demonstrated that CNN-based pose estimation combined with classifier algorithms can effectively identify the possible mode of action of insecticidal compounds, a process that was previously costly and time-consuming with conventional techniques.