Behavior classification: Introducing machine learning approaches for classification of sign-tracking, goal-tracking and beyond

Classifying behaviors in research often relies on predetermined or subjective cutoff values, which can introduce inconsistencies and reduce objectivity. For example, in Pavlovian conditioning studies, rodents display diverse behaviors which can be quantified using the Pavlovian Conditioning Approach (PavCA) Index score. This score is used to categorize subjects as sign-trackers (ST), goal-trackers (GT), or intermediate (IN) groups, but the cutoff values used to distinguish these categories are often arbitrary and inconsistent. The inconsistencies stem from variability in the skewness and kurtosis of score distributions across laboratories, influenced by a range of biological and environmental factors. To address this issue, we explored two approaches to PavCA Index score classification: the k-Means classification and the derivative method. These methods determine cutoff values based on the distribution of PavCA Index scores in the sample, allowing for broader applicability to various types of behavioral data. Our results suggest that these methods, particularly the derivative method based on mean scores from the final days of conditioning, are effective tools for identifying sign-trackers and goal-trackers, especially in relatively small samples. In contrast to existing methods, our approaches provide a standardized classification framework that reflects unique distributions. Furthermore, these methods are adaptable to a researcher's specific needs, accommodating different models and sample sizes. To facilitate implementation, we provide MATLAB code for classifying subjects using both the k-Means classifier and the derivative method.