Explainable automated pain recognition in cats.

Area of Science:

• Veterinary Medicine
• Animal Welfare Science
• Artificial Intelligence in Healthcare

Background:

• Manual pain assessment in cats is subjective and requires expertise.
• Automated pain recognition systems are being developed for various species, including cats.
• Previous studies showed comparable accuracy for deep learning and landmark-based approaches but used homogeneous datasets.

Purpose of the Study:

• To evaluate AI models for classifying pain in a heterogeneous population of cats.
• To compare the generalizability of deep learning versus landmark-based approaches in realistic settings.
• To investigate the explainability of AI pain recognition in cats.

Main Methods:

• Trained AI models using two approaches (deep learning and landmark-based) on facial images from 84 client-owned cats.
• Utilized expert veterinary scoring based on the Glasgow composite measure pain scale and clinical history.
• Analyzed facial features important for AI pain classification.

Main Results:

• The landmark-based approach achieved over 77% accuracy in pain detection.
• The deep learning approach achieved over 65% accuracy.
• Nose and mouth regions were identified as more important for pain classification than ear regions.

Conclusions:

• A landmark-based AI approach demonstrates superior performance for automated pain recognition in diverse cat populations.
• AI models can identify key facial indicators of pain in cats.
• Further research is needed to refine AI for objective animal pain assessment.