ScorER: Exploring Annotation Bias in Vision-Based Neonatal Pain Assessment.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Neonatal Medicine

Background:

• Neonatal pain negatively impacts brain development, necessitating accurate pain assessment.
• Automated pain assessment systems are crucial but face data collection and annotation challenges.
• Vision-based annotation is prone to real-world disturbances, leading to data errors.

Purpose of the Study:

• To introduce Scoring Error Recognition (ScorER), a deep neural network tool for identifying vision-based annotation bias in neonatal pain assessment.
• To enhance the quality and reliability of data used for developing automated neonatal pain assessment systems.

Main Methods:

• Developed a deep neural network-based tool, ScorER.
• Integrated the Grad-CAM algorithm to identify key regions linked to annotation bias.
• Utilized a data-driven approach to analyze the causes of annotation errors.

Main Results:

• ScorER effectively identifies and diagnoses biases in vision-based neonatal pain annotations.
• The Grad-CAM integration pinpoints regions contributing to annotation inaccuracies.
• The study provides a method to improve annotation quality for AI development.

Conclusions:

• ScorER assists caregivers in producing high-quality annotations for neonatal pain.
• This tool supports the creation of large-scale, reliable clinical databases for AI applications.
• Facilitates advancements in AI-driven automated neonatal pain assessment and management.