Can Artificial Intelligence (AI) recognise pain in cats?

27 November 2023

Like it or not, Artificial Intelligence (AI) is increasingly becoming a part of all our lives.

AI brings both opportunities and challenges to veterinary medicine and has the potential for numerous applications, although many clinical directions are yet to be explored and there is currently limited deployment in the clinical setting1,2.

Artificial intelligence refers to the ability of digital systems to perform tasks typically undertaken by humans, such as speech recognition, visual analysis, and decision making1. In recent years, studies have investigated the potential for AI in fields such as thoracic radiography3, diagnostic imaging1 and digital pathology4.

The use of AI in the automation of pain recognition has been relatively unexplored, but a new study published in Scientific Reports has examined whether AI models can classify ‘pain’ or ‘no pain’ in cats, with promising results5. Pain assessment is notoriously difficult in cats and difficulties in recognising and assessing pain are commonly cited in studies of veterinarian’s perception of pain assessment5. Despite the availability of validated manual pain assessment scales for cats, there have been suggestions to develop alternative pain scoring and assessment methods which are more objective, less susceptible to human bias, and applicable to more heterogeneous populations5.

The latest study assessed to what extent a machine can recognise pain in a relatively naturalistic population of cats, and which facial features are most important for the machine in relation to pain recognition5. The study used 84 client-owned cats of different breeds, ages, sex, and varying medical history, presented to the Department of Small Animal Medicine and Surgery of the University of Veterinary Medicine Hannover.

Facial images of cats were captured using a mobile phone. Cats were scored by experienced veterinarians using the validated Glasgow composite measure pain scale (CMPS; which refers to changes in the cat’s behaviour and face) combined with the clinical history of the patients. The images were divided into two classes: ‘pain’ or ‘no pain’. The scoring was used to train AI models using two different approaches: the deep learning (DL) and landmark-based (LDM) approach. Both approaches are based on manual landmark annotations on the cat’s facial alignment. The aligned images are then input to the deep learning models as they are, while the landmark-based approach uses the locations of the 48 landmarks to create multi-vectors according to cat facial regions capturing ears, nose, mouth, and eyes, which forms the final input to the machine learning models.

The results showed that the landmark-based approach performed better, reaching accuracy above 77% in pain detection, and is potentially more suitable for use on more naturalistic cat populations, compared to above 65% accuracy reached by the deep learning approach. A drawback of the landmark-based approach is the time and resource required for landmark annotation, which has to be completed manually. When considering the importance of facial features, the cat’s nose and mouth were more important for machine pain classification features, while features related to the ears were less important.

Whilst the pain detection accuracy achieved by the landmark-based approach (77%) is promising, this does mean that in a clinical setting nearly a quarter of cats could be in pain but it goes unrecognised. Critically, clinical impression that a patient is in pain must always outweigh a binary yes/no or pain/no pain that would be derived from AI, and if ever in doubt, pain relief must be given.

Overall, the study concludes that AI-assisted recognition of pain from cat faces is feasible, and development of such methods will contribute towards accurate automated cat pain recognition in clinical settings. However, it was limited by the relatively small dataset (84 cats), with majority male cats, and the use of still photos which only capture one momentary facial expression (as opposed to video).

Whilst AI evidently brings many opportunities to veterinary medicine, its use raises numerous ethical issues. Coghlan & Quinn (2023) identified nine issues that require ethical consideration when using AI: accuracy and reliability; overdiagnosis; transparency; data security; trust and distrust; autonomy of clients; information overload and skill erosion; responsibility for AI-influenced outcomes; and environmental effects6. They also identified the following principles on guiding use of veterinary AI: nonmaleficence; beneficence; transparency; respect for client autonomy; data privacy; feasibility; accountability; and environmental sustainability6.

As veterinary AI progresses, ethical issues will need to be embedded into education, so veterinarians can effectively protect pets and their human clients, and to ensure the safe introduction of AI technologies into clinical practice6.


BSAVA library collection on pain management


1Hennessey E, DiFazio M, Hennessey R & Cassel N (2022) Artificial intelligence in veterinary diagnostic imaging: A literature review. Veterinary Radiology & Ultrasound. 63 (1), 851-870.

2Basran PS & Appleby (2022) The unmet potential of artificial intelligence in veterinary medicine. Am J Vet Res. 83(5), 385-392. doi: 10.2460/ajvr.22.03.0038.

3Li S, Wang Z, Visser LC, Wisner ER & Cheng H (2020) Pilot study: Application of artificial intelligence for detecting left atrial enlargement on canine thoracic radiographs. Veterinary Radiology & Ultrasound. 61 (6), 611-618.

4Zuraw A & Aeffner F (2022) Whole-slide imaging, tissue image analysis, and artificial intelligence in veterinary pathology: An updated introduction and review. Veterinary Pathology. 59(1), 6-25. doi:10.1177/03009858211040484

5Feighelstein M, Henze L, Meller S. et al. (2023) Explainable automated pain recognition in cats. Scientific Reports. 13, 8973.

6Coghlan S & Quinn T (2023) Ethics of using artificial intelligence (AI) in veterinary medicine. AI & Society.