Pain Action Unit Detection in Critically Ill Patients.

Area of Science:

• Medical Technology
• Artificial Intelligence
• Critical Care Medicine

Background:

• Current pain assessment in ICUs relies on subjective self-reports or limited behavioral observations.
• Assessing pain in non-verbal patients is difficult, subjective, and increases staff workload.
• Previous facial Action Unit (AU) detection methods for pain were limited to controlled environments.

Purpose of the Study:

• To develop and validate an automated system for detecting facial pain expressions (AUs) in critically ill patients within an uncontrolled ICU setting.
• To address the limitations of existing pain assessment methods in intensive care units.

Main Methods:

• Collected and annotated the novel "Pain-ICU" dataset, comprising 55,085 images of critically ill adult patients.
• Evaluated OpenFace, an open-source AU detection tool, and a custom-trained AU R-CNN model on the Pain-ICU dataset.
• Accounted for challenging variables in ICU environments, including assisted breathing devices and variable lighting.

Main Results:

• OpenFace demonstrated poor performance in detecting AUs in the ICU setting, with an F1-score of 0.42.
• The AU R-CNN model, trained on the Pain-ICU dataset, achieved a significantly higher average F1-score of 0.77.
• Demonstrated the feasibility of detecting facial pain AUs in uncontrolled ICU environments.

Conclusions:

• Automated detection of facial pain AUs is feasible in challenging, real-world intensive care unit settings.
• The developed AU R-CNN model offers a more accurate and objective method for pain assessment in non-verbal ICU patients.
• This technology has the potential to improve pain management and patient care in critical care.