Assessment of Pain Intensity Using Deep Learning Models in Non-Communicative Intensive Care Patients.

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Signal Processing
• Clinical Informatics

Background:

• Pain assessment in intensive care units (ICUs) is challenging for non-communicating patients.
• Conventional pain assessment tools are subjective and prone to observer bias.
• Deep learning facial analysis offers objective quantification of pain-related behaviors.

Purpose of the Study:

• To evaluate deep learning models for pain severity classification in non-communicative adult ICU patients.
• To assess the feasibility and diagnostic accuracy of AI-driven facial analysis for pain.

Main Methods:

• Utilized DenseNet-169 for feature extraction and principal component analysis for dimensionality reduction.
• Classified pain severity using Support Vector Machine (SVM), Random Forest, and K-Nearest Neighbors (KNN).
• Facial images were annotated by a multidisciplinary team (intensivist, nurses, pain specialist).

Main Results:

• The SVM model achieved 96.9% accuracy and 0.994 AUC, with high sensitivity for severe pain.
• KNN demonstrated superior performance for moderate pain detection.
• Low inter-rater agreement (k=0.16) among experts highlighted subjective variability in manual assessment.

Conclusions:

• Deep learning facial analysis provides a valid, reproducible, and standardized method for pain assessment in non-verbal ICU patients.
• AI minimizes inter-observer variability, offering objective decision support for critical care nurses.
• This technology standardizes pain management and enhances patient safety by reducing assessment bias.