Abstract:
Clinical text summarization is a critical yet challenging task due to the unstructured
and heterogeneous nature of electronic health records (EHRs). Discharge
summaries, in particular, often contain lengthy and complex narratives, making
it difficult to efficiently extract key clinical insights. Recent advancements in generative
large language models (LLMs) have shown great promise in automating
summarization tasks; however, their optimal application in clinical contexts remains
underexplored, representing a significant research gap. To address this gap,
the present study investigates an adaptive prompting approach for discharge summary
summarization using different few-shot learning settings, ranging from oneshot
to six-shot, with several state-of-the-art generative LLMs, including GPT-5,
LLaMA, Mistral, Gemini, and T5. The performance of these models was evaluated
using multiple metrics—such as F1 score, ROUGE, and BERTScore—alongside
complementary measures like BLEU, BLEURT, and METEOR, to ensure a comprehensive
assessment. The experimental results revealed that GPT-5 achieved
superior performance across all evaluation metrics in the five-shot configuration,
reaching an average score of 99%. This outcome underscores the strong effectiveness
of five-shot learning for clinical summarization tasks. These findings provide
valuable insights for healthcare researchers, informatics practitioners, and policymakers
aiming to integrate LLMs into clinical workflows to support efficient
information retrieval and informed clinical decision-making.
