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

Neonatal pain can impair brain development, highlighting the critical clinical importance of accurate pain assessment. In recent years, automated neonatal pain assessment systems have received substantial research attention. However, their advancement is hindered by challenges in data collection and annotation. Vision-based neonatal pain annotation is convenient but susceptible to real-world disturbances that may result in information loss and incorrect annotations. In this study, we propose a deep neural network-based tool, Scoring Error Recognition (ScorER), which aims to automatic identification and intelligent diagnosis of vision-based annotation bias. By integrating the Grad-CAM algorithm, we systematically identify key regions associated with annotation bias and employ a data-driven approach to investigate their underlying causes. This research proposes a novel approach to improve annotation quality and provides new insights for advancing automated neonatal pain assessment systems in future studies.Clinical relevance- ScorER can assist caregivers in achieving high-quality vision-based neonatal pain annotations, supporting the development of large-scale, high-quality clinical neonatal pain databases for real-world applications. This facilitates AI-driven automated neonatal pain assessment research and system development for precise neonatal pain management.