Embedding-Enhanced Patient Clustering for Customized Medical Report Summarization using LLMs

This paper presents a novel approach to enhancing medical report summarization by integrating Large Language Models (LLMs) with patient data embedding and clustering techniques. Traditional medical text summarization often overlooks the unique contextual needs of individual patients, resulting in suboptimal summaries. To address this limitation, we propose leveraging embedding techniques, such as t-SNE, to extract relevant lower-dimensional features from patient clinical data, followed by clustering to identify meaningful patient groups. Next, we apply an LLM to generate customized summaries from medical reports for each cluster, ensuring alignment with specific medical contexts. These summaries undergo an iterative refinement process with healthcare professionals to enhance their clinical relevance and accuracy. The refined summaries are then used as labels for the original reports to fine-tune the LLM, improving its predictive capabilities. In practice, due to limited access to real expert feedback and refinement, we implemented our approach up to the stage of generating summaries, leaving further refinement and fine-tuning of the LLM for future work. We evaluate the effectiveness of our methodology using a synthetic dataset, demonstrating its feasibility and potential to advance LLM-based medical text summarization. Furthermore, we identified a meaningful relationship between the quality of the generated clusters, in terms of their structure, and the relevance of the summaries produced, as compared to the true summaries. This relationship was observed based on the bestperforming embedding-clustering model. As an impact, this approach enhances both the efficiency of healthcare providers and the clarity of medical reports, potentially reducing cognitive workload while maintaining high clinical fidelity.