Kenneth Craig2, Jeannie S. Huang1, Xiaojing Xu1, Damaris Diaz1, Virginia de Sa1
1) United States 2) Canada
Background: We have demonstrated the validity of facial expression technology in assessing presence v. absence and severity of clinically significant post-operative pain in youth undergoing laparoscopic appendectomy (LA). Our primary objective was to determine the ability of automated facial expression recognition technology to correctly classify real v. masked v. faked v. no pain facial expression in youth.
Methods: 48 youth (5 – 17 yr) were studied within 48 hrs of and 3 weeks after LA. At each session, participants were video recorded during a 10-sec pain stimulus (abdominal pressure) and provided NRS pain level reports, 0 to 10. Youth also either suppressed pain expression or faked worst pain ever during pain stimulus. Videos were classified as real v. masked v. faked v. no pain using NRS scores >=4 v. <=3). We trained an automated machine learning model based on pain-related Facial Action Coding System action units (extracted using Emotient software) to detect real v. masked v. faked v. no pain in a subset of videos and then used cross-validation to determine model performance across the entire collected set of videos.
Results: Machine performance demonstrated greater accuracy at distinguishing real v. masked v. faked v. no pain states relative to human judgments in prior research.
Conclusions: Our facial expression machine model demonstrated promising ability to distinguish various clinical pain states in youth following surgery. The spontaneous expression of pain can be differentiated from facial expressions voluntarily suppressed or dissembled.
Boerner, K.E., Chambers, C.T., Craig, K.D., Pillai Riddell, R.R., Parker, J.A. Caregiver accuracy in detecting deception in facial expressions of pain in children. Pain 2013; 143: 525-533
Sikka, K., Ahmed, A.A., Diaz, D., Goodwin, M.S., Craig, KI.D., Bartlett, M.S., Huang, J.S. Automated assessment of children’s postoperative pain using computer vision. Pediatrics 2015; 136: e124-131.
Acknowledgement: Funding from the USNIH NINR grant 13500
