Explanation-Guided Medical Named Entity Recognition with Stability and Boundary Awareness for Atopic Dermatitis

Researchers propose an explanation-guided framework for medical named entity recognition (NER) in Chinese atopic dermatitis clinical texts, using stability and boundary-aware constraints to improve model reliability and interpretability. The method combines perturbation-based analysis with adaptive fusion of local and global explanations, achieving performance gains across multiple NER models while enhancing explanation robustness for clinical decision support.

This research addresses a critical gap in medical AI interpretability by developing a framework that makes named entity recognition systems more transparent and reliable for clinical applications. Traditional NER models often function as black boxes, making clinicians hesitant to trust their outputs for patient care decisions. By integrating explanation-guided learning with stability constraints, the framework ensures that model decisions are not only accurate but also understandable to medical professionals.

The significance lies in the methodology's dual focus on performance and interpretability. Most NER systems prioritize accuracy metrics while treating explainability as secondary, but healthcare applications demand both. The perturbation-based analysis and boundary-aware constraints directly address common failure modes in medical text processing, where entity boundaries often determine clinical relevance. For atopic dermatitis specifically, precise entity extraction from clinical notes enables better symptom tracking and treatment monitoring.

For healthcare AI developers and clinical institutions, this framework provides a practical pathway to deploy NER systems with higher confidence. The adaptive fusion strategy that combines multiple explanation methods represents a generalizable approach applicable beyond dermatology to other medical specialties. The work demonstrates that explainability and robustness need not be sacrificed for performance—the framework achieves improvements across both dimensions simultaneously.

Future applications may extend this methodology to multilingual medical texts and additional clinical domains, particularly in languages like Chinese where NER challenges are more pronounced. As healthcare systems increasingly adopt AI-assisted documentation and analysis, frameworks that prioritize explanation reliability become essential infrastructure rather than optional enhancements.

• →Explanation-guided learning improves both NER accuracy and interpretability in clinical medical texts, addressing trust barriers in AI-assisted healthcare.
• →Adaptive fusion of local and global explanations provides more stable and boundary-aware token-level interpretations than single explanation methods.
• →The framework's stability and boundary-aware constraints ensure reliable entity extraction for downstream clinical decision-making applications.
• →Methodology is generalizable across NER models and medical specialties, providing practical value beyond atopic dermatitis.
• →Integration of explainability constraints during training achieves performance gains without sacrificing model transparency or robustness.