Machine learning-based automated phenotyping of inflammatory nocifensive behavior in mice.

Abstract

The discovery and development of new and potentially nonaddictive pain therapeutics requires rapid, yet clinically relevant assays of nociception in preclinical models. A reliable and scalable automated scoring system for nocifensive behavior of mice in the formalin assay would dramatically lower the time and labor costs associated with experiments and reduce experimental variability. Here, we present a method that exploits machine learning techniques for video recordings that consists of three components: key point detection, per frame feature extraction using these key points, and classification of behavior using the GentleBoost algorithm. This approach to automation is flexible as different model classifiers or key points can be used with only small losses in accuracy. The adopted system identified the behavior of licking/biting of the hind paw with an accuracy that was comparable to a human observer (98% agreement) over 111 different short videos (total 284 min) at a resolution of 1 s. To test the system over longer experimental conditions, the responses of two inbred strains, C57BL/6NJ and C57BL/6J, were recorded over 90 min post formalin challenge. The automated system easily scored over 80 h of video and revealed strain differences in both response timing and amplitude. This machine learning scoring system provides the required accuracy, consistency, and ease of use that could make the formalin assay a feasible choice for large-scale genetic studies.