Construction and validation of a pain facial expressions dataset for critically ill children

Area of Science:

• Pediatric critical care
• Medical imaging and machine learning
• Pain assessment

Background:

• Automatic pain assessment for non-communicative children is crucial but hindered by limited training data.
• Existing datasets lack sufficient scale and specificity for diverse pediatric populations.

Purpose of the Study:

• To create a large-scale, high-quality dataset of pain facial expressions for Chinese critically ill children.
• To evaluate the utility of this dataset using deep learning models for pain assessment.
• To establish a benchmark for automatic pain detection in pediatric intensive care.

Main Methods:

• Collected 119 pain expression videos and 6951 images from 53 critically ill children in intensive care units.
• Triple-labeled all data independently across five pain intensity levels.
• Developed and evaluated deep learning models using the "pain facial expression of critically ill children" (PFECIC) dataset.

Main Results:

• The PFECIC dataset demonstrated strong performance with deep learning models, achieving 88.3% accuracy, 88.3% precision, 88.7% recall, and an 88.5% F1-score.
• The model exhibited a low false-positive rate of 3.0%.
• Comparative analysis showed PFECIC outperformed the COPE dataset in accuracy, validity, and comprehensiveness.

Conclusions:

• The PFECIC dataset is a valuable resource for developing and validating automatic pain assessment tools for critically ill children.
• Deep learning models trained on PFECIC show significant potential for accurate and reliable pain detection.
• This work addresses the critical need for specialized datasets in pediatric pain research.