Computer vision for pain detection during procedural sedation.
Authors
Affiliations (9)
Affiliations (9)
- KITE Research Institute, Toronto Rehabilitation Institute - University Health Network, Toronto, ON, Canada. [email protected].
- Department of Computer Science, University of Toronto, Toronto, ON, Canada. [email protected].
- Lawrence S. Bloomberg Faculty of Nursing, University of Toronto, Toronto, ON, Canada.
- Division of Vascular and Interventional Radiology, Department of Medical Imaging, University of Toronto, Toronto, ON, Canada.
- KITE Research Institute, Toronto Rehabilitation Institute - University Health Network, Toronto, ON, Canada.
- Department of Computer Science, University of Toronto, Toronto, ON, Canada.
- Department of Biomedical Engineering, University of Toronto, Toronto, ON, Canada.
- Vector Institute for Artificial Intelligence, Toronto, ON, Canada.
- School of Nursing, Queensland University of Technology (QUT), Brisbane, QLD, Australia.
Abstract
Continual automated pain detection from facial expressions using computer vision during procedural sedation could optimize sedation dose titration and minimize pain. A prospective observational study was conducted. Participants' faces were recorded during interventional radiology procedures performed with procedural sedation. Simultaneous pain assessments were made by a nurse using a sedation state assessment scale. Videos from 70 participants were used to train and evaluate a pain detection model using the Swin Transformer architecture. The model demonstrated an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.79 and an Area Under the Precision-Recall Curve (AP) of 0.53. The model reliably identified pain events for many participants, with sharp increases in predicted probabilities closely aligning with actual pain occurrences. Exploratory subgroup analyses suggested variability in performance across Fitzpatrick skin tone categories (highest in categories 3-4), but subgroup sizes were limited and these findings require validation in larger, independent cohorts. Automated pain detection systems for procedural sedation using this model would require a high threshold that minimizes false-positive alerts for pain to limit the risk of alarm fatigue. Alternatively, user interfaces that display predicted probabilities over time without alerts may be preferable until further refinements are made to enhance model performance and fairness across the diverse population of patients who receive procedural sedation.