TT-XAI: Trustworthy Clinical Text Explanations via Keyword Distillation and LLM Reasoning

Keyword distillation improves both accuracy and explainability: The paper demonstrates that condensing lengthy electronic health records into domain-aware keywords simultaneously enhanced BERT classifier performance and made explanations more faithful to the model's actual decision-making process, suggesting that simpler representations can lead to more trustworthy AI systems.

Human experts prefer AI explanations guided by structured information: A blinded study with domain experts consistently favored explanations generated when the AI system was guided by distilled keywords rather than raw text, indicating that how we structure information for AI systems directly impacts whether clinicians find their reasoning trustworthy and useful.

Multiple evaluation methods should validate explanation quality: The research used three different approaches to assess explanation quality—deletion-based fidelity metrics, automated LLM scoring, and human expert review—and all three agreed on the superiority of the keyword-augmented method, suggesting that responsible AI evaluation benefits from diverse validation approaches rather than relying on a single metric.

Lightweight frameworks can make clinical AI more auditable: TT-XAI achieves its improvements through a relatively simple approach (keyword extraction plus guided prompts) rather than complex black-box methods, offering a practical pathway for healthcare organizations to deploy AI systems that clinicians can actually understand and audit.