Abstract
Quantification and parametrization of movement in animal models is widely used in behavioral paradigms. In particular, free movement of an animal in controlled conditions (e.g., the open field paradigm) is used as a proxy for indices of baseline and drug-induced behavioural changes. However, the analysis of this is often time- and labour-intensive and existing algorithms do not always classify the behaviour correctly.
Here, we propose a new approach to quantify behaviour in an unconstrained environment: searching for frequent patterns (k-motifs) in the time series representing position of the subject over time. Validation of this method was performed using subchronic quinpirole-induced changes in open field experiment behaviors in rodents. Analysis of this data was performed using k-motifs as features to better classify subjects into experimental groups on the basis of behavior in the open field. Our classifier using k-motifs gives as high as 94% accuracy in classifying repetitive behaviour versus controls which is a substantial improvement compared to currently available methods including using standard feature definitions (depending on the choice of feature set and classification strategy, accuracy up to 88%). Furthermore, vizualization of the movement / time patterns is highly predictive of these behaviours. By using machine learning to create features in a data driven fashion, this can be applied to general behavioural analysis across experimental paradigms beyond the open field.