Flying In-formation: A computational method for the classification of host seeking mosquito flight patterns using path segmentation and unsupervised machine learning

Abstract

The rational design of effective vector control tools requires detailed knowledge of vector behaviour. Yet, behavioural observations, interpretations, evaluations and definitions by even the most experienced researcher are constrained by subjectivity and perceptual limits. Seeking an objective alternative to expertise, we developed and tested an unsupervised method for the automatic identification of video-tracked mosquito flight behaviour. This method unites path-segmentation and unsupervised machine learning in an innovative workflow and is implemented using a combination of R and python. The workflow (1) records movement trajectories; (2) applies path-segmentation; (3) clusters path segments using unsupervised learning; and (4) interprets results. Analysis of the flight patterns of An. gambiae s.s., responding to human-baited insecticide-treated bednets (ITNs), by the new method identified four distinct behaviour modes: with swooping and approaching modes predominant at ITNs; increased walking behaviours at untreated nets; similar rates of reacting at both nets; and higher overall activity at treated nets. The methods validity was tested by comparing these findings with those from a similar setting using an expertise-based method. The level of correspondence found between the studies validated the accuracy of the new method. While researcher-defined behaviours are inherently subjective, and prone to corollary shortcomings, the new approachs mathematical method is objective, automatic, repeatable and a validated alternative for analysing complex vector behaviour. This method provides a novel and adaptable analytical tool and is freely available to vector biologists, ethologists and behavioural ecologists. Author summaryVector control targets the insects and arachnids that transmit 1 in every 6 communicable diseases worldwide. Since the effectiveness of many vector control tools depends on exploiting or changing vector behaviour, a firm understanding of this behaviour is required to maximise the impact of existing tools and design new interventions. However, current methods for identifying such behaviours are based primarily on expert knowledge, which can be inefficient, difficult to scale and limited by perceptual abilities. To overcome this, we present, detail and validate a new method for categorising vector behaviour. This method combines existing path segmentation and unsupervised machine learning algorithms to identify changes in vector movement trajectories and classify behaviours. The accuracy of the new method is demonstrated by replicating existing, expert-derived, findings covering the behaviour of host-seeking mosquitos around insecticide treated bednets, compared to nets without insecticide. As the method found the same changes in mosquito activity as previous research, it is said to be validated. The new method is significant, as it improves the analytical capabilities of biologists working to reduce the burden of vector-borne diseases, such as malaria, through an understanding of behaviour.