Automated acute pain assessment in brachycephalic cats with ocular pain using the Feline Grimace Scale.

OBJECTIVES

This study aimed to evaluate the feasibility and limitations of an AI-powered Feline Grimace Scale (FGS) pipeline for acute pain assessment in brachycephalic cats with ocular pain.

METHODS

A total of 200 facial images of brachycephalic cats with ocular pain were used for model training and evaluation. A landmark localization model was used to predict facial landmarks. Pain prediction was performed with regression models from a machine-learning library (XGBoost; XGBoost Developers); 5 independent submodels were trained, each corresponding to 1 of the 5 FGS action units-ear position, orbital tightening, muzzle tension, whiskers change, and head position. Three configurations were applied to input variables: without muzzle tension and whiskers change, without whiskers change and head position, and all geometric descriptors retained. The landmark localization performance was assessed via normalized root mean square error (NRMSE), whereas pain prediction was assessed via mean absolute error, Pearson correlation, and R2 values.

RESULTS

Facial landmark detection showed reduced accuracy in brachycephalic cats (NRMSE, 19.96%) compared with previous reports in predominantly mesocephalic cats (NRMSE, 16.76%). Pain prediction models demonstrated poor generalizability, showing the best results for without muzzle tension and whiskers change (mean absolute error, 0.168; R2, 0.115; Pearson correlation, 0.374).

CONCLUSIONS

Automated landmark detection and pain prediction using an AI-powered FGS pipeline is possible in brachycephalic cats. However, cautious interpretation is needed due to limited annotated data, which precludes robust predictive performance estimates.

CLINICAL RELEVANCE: Conformation-specific datasets are required before this AI-powered FGS pipeline can be used for pain assessment in brachycephalic cats.