Deep Learning-Based Pain Classifier Based on the Facial Expression in Critically Ill Patients.

Area of Science:

• Medical Artificial Intelligence
• Critical Care Medicine
• Computer Vision

Background:

• Pain assessment in critically ill patients is crucial but lacks objective tools.
• Facial expressions are key indicators of pain in non-communicative patients.
• Existing methods for pain assessment in this population are often subjective and unreliable.

Purpose of the Study:

• To develop and validate a deep learning-based automated pain classifier using facial expressions.
• To establish both image- and video-based models for pain assessment.
• To evaluate the performance of these classifiers in a real-world clinical setting.

Main Methods:

• Prospective study involving critically ill patients (2020-2021).
• Video recordings of patients with labeled pain scores (relaxed, tense, grimacing).
• Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) models (Resnet34, VGG16, InceptionV1) were utilized.

Main Results:

• Video-based classifiers achieved high accuracy (e.g., ~0.88 for 0 vs. 2 pain score).
• Image-based classifiers also demonstrated significant accuracy (e.g., ~0.86 for 0 vs. 2 pain score).
• Classifiers maintained high performance even when tested on new patients without prior reference.

Conclusions:

• Deep learning offers a practical solution for automated pain assessment in critically ill patients.
• The developed AI tool shows potential for objective and reliable pain monitoring.
• Further validation studies are recommended to confirm these findings in diverse clinical settings.