Background & Aims
Prior research on pain recognition and bias reveal that humans demonstrate bias when appraising the pain experiences of others 1. The appraisee’s racial identity and biological sex may also influence the appraiser’s assessment 1,2. In our study, which explored biobehavioral and sociocultural influences in pain expression and assessment, we attempted to decrease these biases in pain assessment by utilizing Artificial Intelligence (AI) instead of humans. However, AI may also be biased as it was made by humans and historically trained with homogenous samples (e.g., similar demographic profiles). In the current study, we aimed to determine whether bias exists in AI software in its ability to detect faces based on the race or sex of the face or the general diversity present across the dataset.
Methods
Participants from 2 different studies displayed pain-evoked facial expressions. In study 1, 166 individuals at the NIH received thermal stimuli in 10 second trials while we recorded their facial expressions. We compared this dataset with the Denver Pain Authenticity Stimulus Set (DPASS)3 (N=103) in which individuals received 3-s pressure for 1 trial. We used a commercially-available AI software, iMotions5, to track facial movement coordinates, assessing muscle activation intensity at every time point. For each trial and participant, we identified the proportion of time that the software did not recognize facial features. Data points with detected features were labeled “usable”. We concatenated Average Mean Percentage Usability (AMPU) scores across all participants and trials. Our analysis focuses on responses in White, Black, and multiracial participants across both datasets. A 3-way ANOVA was used to determine whether AMPU varied as a function of Study, Race, Sex and all interactions.
Results
We observed a significant main effect of Study (p= 0.01), but there were no significant main effects of Sex, Race, or interactions with sex or race (all p’s > 0.1). Average Mean Percentage Usability was higher in the DPASS study (M = 98.12, SD = 11.94) relative to the study we conducted (M = 92.66, SD = 14.78), though AMPU was above 90% across both studies.
Conclusions
Our results demonstrate that AI software is a viable tool for feature detection without bias related to race or sex when considering usability. However, these results also suggest the need for further investigation into other factors that may affect AI performance. Future studies should explore these dimensions to better understand and enhance the robustness of AI application in facial feature detection.
References
1. Hoffman, K. M., Trawalter, S., Axt, J. R., & Oliver, M. N. (2016). PNAS, 113(16), 4296–4301.
2. Mende-Siedlecki, P., Qu-Lee, J., Backer, R., & Van Bavel, J. J. (2019). JEP-General, 148(5), 863–889.
3. Lloyd, E. P., Summers, K.M., Gunderson, C., Weesner, R. E., ten Brinke, L.,Hugenberg, K., & McConnell, A. R., (under review). Denver pain authenticity stimulus set (D-PASS). Manuscript Under Review.
4. Mogil, J. Sex differences in pain and pain inhibition: multiple explanations of a controversial phenomenon. Nat Rev Neurosci 13, 859–866 (2012). https://doi.org/10.1038/nrn3360
5. iMotions (9.3), iMotions A/S, Copenhagen, Denmark, (2022).
6. Images Created with BioRender.com
Presenting Author
Ruth Mosunmade
Poster Authors
Ruth Mosunmade
BA
National Center for Complementary and Integrative Health, National Institutes of Health
Lead Author
Yili Zhao
PhD
National Center for Complementary and Integrative Health, National Institutes of Health
Lead Author
Troy Dildine
PhD
Department of Anesthesiology, Perioperative and Pain Medicine,Stanford University School of Medicine
Lead Author
Lauren Atlas
PhD
National Center for Complementary and Integrative Health, National Institutes of Health
Lead Author
Topics
- Other