Explainable Artificial Intelligence for Clinical Text Classification: A Scoping Review of Methods and Applications

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing for Clinical Data

Background:

• Machine learning (AI) systems in healthcare often function as "black boxes," limiting interpretability and clinical trust.
• Transparency is crucial for adopting AI in clinical settings, especially when analyzing unstructured electronic health record (EHR) data.
• Clinical text classification tasks require interpretable AI models to ensure reliable decision-making.

Purpose of the Study:

• To conduct a scoping review of explainable artificial intelligence (XAI) approaches applied to clinical text classification.
• To analyze the methods used for interpreting AI models processing unstructured EHR data.
• To identify trends and performance of XAI techniques in healthcare AI applications.

Main Methods:

• Systematic literature search and analysis of 22 studies focusing on XAI in clinical text classification.
• Categorization of studies based on classification tasks (binary, multi-class, multi-label) and AI model architectures.
• Review of both model-specific (e.g., attention mechanisms) and model-agnostic (e.g., SHAP, LIME) XAI techniques.

Main Results:

• Transformer-based models demonstrated superior performance in clinical text classification tasks.
• Attention mechanisms were frequently employed as model-specific XAI methods within deep learning models.
• Model-agnostic methods like SHAP and LIME offered versatile interpretability across various AI architectures.

Conclusions:

• XAI methods significantly enhance the transparency and trustworthiness of AI systems used in clinical text analysis.
• The study highlights the effectiveness of both specialized and general XAI approaches for EHR data.
• Further research is necessary to validate the clinical utility and reliability of AI explanations in healthcare.