Adults' Pain Recognition via Facial Expressions Using CNN-Based AU Detection

The identification of pain expression by facial analysis is a difficult problem with few effective solutions. Most of the existing facial expression recognition methods are based on the detection of different Action Units (AU) to identify simple expressions such as anger, joy or neutral. In this study, we propose using CNNs pre-trained on previously developed models for pain detection, such as the models included in PainChek, a point-of-care mobile application that uses automated facial assessment and analysis in the assessment of procedural pain in infants, to combine nine AUs identified in video images from the MORPH database without specific pain and from a proprietary database with joint pain. This work also analyses the variability of the AU detection method applied to images of individuals of different age, race and gender. For the database without specific pain, the mean amount of AUs considered as not present was 92.38% and 7.62% for present, obtaining the worst results for the identification of AU6 and AU12 with 83.67% of present and 81.20% of not present. For the joint pain database with ground truth labelled by AU and intensity, the mean accuracy for the 9 AUs was 84.98% with a sensitivity of up to 92% percent for AU25 and AU43.

Automated summary

The identification of pain expression by facial analysis is challenging. This study proposes using CNNs pre-trained on pain detection models to combine different AUs for pain recognition. The method shows variability in AU detection.