AI-powered rodent behavior recognition model for pain research

Pain assessment in rodent models relies on behavioral observation, yet manual scoring is labor-intensive and variable; comprehensive automated systems capable of simultaneously quantifying multiple pain-related behaviors in rats remain limited.

We developed and evaluated an artificial intelligence (AI)-powered behavioral recognition model that identifies and quantifies four pain-related behaviors in rats. Video recordings of rats were obtained, and 264 clips were annotated into four behavioral categories: body grooming, face grooming, freezing, and exploratory behavior. These annotations were used to train and test an AI model. Performance was further assessed on five long, unseen continuous videos (6-8 min each) to evaluate generalizability and real-world applicability.

The AI model classified the four target behaviors, providing continuous temporal recognition in 3-second time windows and quantifying the frequency and duration of each behavior. The AI model achieved an overall accuracy of 91.7% on the held-out balanced short-clip test set. When further evaluated on five longer unseen videos using an overlapping 3-second window inference strategy with majority voting, the model achieved an overall accuracy of 73.2% with manual behavioral annotations across 776 matched 3-second timeframes. Class-specific accuracy was highest for body grooming (84.8%) and exploratory behavior (75.4%), followed by freezing (63.3%) and face grooming (52.4%).

Compared with pose-based or grimace-focused approaches, the present framework directly classifies and continuously quantifies four spontaneous rat behaviors from video.

The AI-powered automated behavioral recognition system may offer an objective, reproducible, and scalable approach to behavioral assessment in rats, and may enhance the translational relevance of preclinical pain research and analgesic development.