Emotion-Aware Knowledge Distillation: A Patient-Centric Framework for Biomedical Dialogue Summarization

Abstract While clinical dialogue summarization systems can alleviate documentation burdens, current approaches focus solely on informational content like symptoms and diagnoses. These systems often overlook the patient’s emotional signals, such as fear, anxiety, or anger. Although Large Language Models (LLMs) capture these nuances, high computational costs and data privacy risks restrict their real-time clinical integration. To overcome these obstacles, this study proposes an innovative and hybrid framework based on Emotion-Aware Knowledge Distillation. In this study, the empathy and reasoning capabilities of the Gemini teacher model are distilled via synthetic data generation strategies and transferred to a resource-efficient, compact proposed student model (fine-tuned BioBART-v2). The developed model improved the signal-to-noise ratio in clinical dialogues through a patient-centric filtering strategy and reduced politeness bias from 44.49% to 0.20%. Experimental results demonstrate that proposed approach achieves a performance increase of 26-28 points in ROUGE and METEOR metrics compared to reference models. Beyond automated metrics, a clinical expert evaluation conducted on 200 clinical dialogues confirmed the model’s real-world applicability, verifying an overall clinical accuracy of 89.00%. As a result of evaluations conducted across three distinct categories, 59.50% of the summaries were classified as perfect, 29.50% as incomplete but accurate, and 11.00% as requiring revision. These results demonstrate the framework’s applicability and safety profile in clinical settings. Furthermore, the system is equipped with a control mechanism that adapts the summary’s tone based on the detected emotional state. The findings reveal that high-accuracy, empathetic, and patient-centric clinical decision support systems can be developed even in resource-constrained environments.