Finding a Husband: Using Explainable AI to Define Male Mosquito Flight Differences

Simple Summary Mosquitoes are vectors of some of the world's deadliest diseases. A wide range of intervention tools are available, but the majority of these depend on insecticides and must be updated frequently in response to the rise of insecticide resistance in the targeted mosquito populations. The behaviour of mosquitoes is an important factor to add to their ongoing understanding as a vector for disease. The method presented within this paper is able to identify key behavioural differences between male, female, and in copula mosquito pairs by analysing their flight tracks. To identify these differences, we developed a framework that extracts features from track segments. Track segments are used to unify durations and each segment is individually classified as either male or non-male by a machine learning model. The segments for each track are combined to return an overall prediction of the class of the track. This approach is one of the first applications of machine learning to mosquito trajectory analysis. The framework can be extended to analyse differences between other classes of trajectories. Mosquito-borne diseases account for around one million deaths annually. There is a constant need for novel intervention mechanisms to mitigate transmission, especially as current insecticidal methods become less effective with the rise of insecticide resistance among mosquito populations. Previously, we used a near infra-red tracking system to describe the behaviour of mosquitoes at a human-occupied bed net, work that eventually led to an entirely novel bed net design. Advancing that approach, here we report on the use of trajectory analysis of a mosquito flight, using machine learning methods. This largely unexplored application has significant potential for providing useful insights into the behaviour of mosquitoes and other insects. In this work, a novel methodology applies anomaly detection to distinguish male mosquito tracks from females and couples. The proposed pipeline uses new feature engineering techniques and splits each track into segments such that detailed flight behaviour differences influence the classifier rather than the experimental constraints such as the field of view of the tracking system. Each segment is individually classified and the outcomes are combined to classify whole tracks. By interpreting the model using SHAP values, the features of flight that contribute to the differences between sexes are found and are explained by expert opinion. This methodology was tested using 3D tracks generated from mosquito mating swarms in the field and obtained a balanced accuracy of 64.5% and an ROC AUC score of 68.4%. Such a system can be used in a wide variety of trajectory domains to detect and analyse the behaviours of different classes, e.g., sex, strain, and species. The results of this study can support genetic mosquito control interventions for which mating represents a key event for their success.

Automated summary

This paper presents a method for identifying key behavioral differences between male, female, and in copula mosquito pairs by analyzing their flight tracks. The approach uses machine learning to classify track segments as either male or non-male and combines the segment classifications to classify the overall track. This methodology provides valuable insights into the behavior of mosquitoes and other insects.