Tracking changes in behavioural dynamics using prediction error.

Tom Lorimer,Vitul Agarwal,Erik Saberski,Rachel Goodridge,George Sugihara,Antonia K Bock,Scott A Rifkin,Giorgio F Gilestro

doi:10.1371/journal.pone.0251053

Tom Lorimer, Vitul Agarwal + Show 6 more

Open Access

https://doi.org/10.1371/journal.pone.0251053

Copy DOI

Abstract

Automated analysis of video can now generate extensive time series of pose and motion in freely-moving organisms. This requires new quantitative tools to characterise behavioural dynamics. For the model roundworm Caenorhabditis elegans, body pose can be accurately quantified from video as coordinates in a single low-dimensional space. We focus on this well-established case as an illustrative example and propose a method to reveal subtle variations in behaviour at high time resolution. Our data-driven method, based on empirical dynamic modeling, quantifies behavioural change as prediction error with respect to a time-delay-embedded ‘attractor’ of behavioural dynamics. Because this attractor is constructed from a user-specified reference data set, the approach can be tailored to specific behaviours of interest at the individual or group level. We validate the approach by detecting small changes in the movement dynamics of C. elegans at the initiation and completion of delta turns. We then examine an escape response initiated by an aversive stimulus and find that the method can track return to baseline behaviour in individual worms and reveal variations in the escape response between worms. We suggest that this general approach—defining dynamic behaviours using reference attractors and quantifying dynamic changes using prediction error—may be of broad interest and relevance to behavioural researchers working with video-derived time series.

Highlights

Behaviour mediates individual interaction with the outside world
Average prediction error remains stable as a function of both E and θ, over a wide range of parameter settings, for almost all worms of each type (Fig 2)
Comparing predictive output at a single time point focuses this information to single-sample resolution. In order for such fine scale prediction to be meaningful, it has to be able to exploit relevant information globally in time—achieved here using the attractor. This data-driven focusing of relevant information across time scales is a natural advantage of the empirical dynamic modeling (EDM) approach

Summary

Introduction

Behaviour mediates individual interaction with the outside world. A systematic description of behaviour is key to linking dynamic internal (e.g. neural) states, with external biotic and abiotic conditions in natural situations. Even in cases where behaviour can be observed, finding a simple quantitative dynamic description is challenging. Motion and pose of freely-moving organisms can be automatically tracked and quantified using computer vision [1].

Methods

Results

Conclusion