GoT-HCS: Graph-of-Thoughts Enhanced Hierarchical Reasoning for Automated Clinical Hospital Course Summarization

Abstract

Brief Hospital Course (BHC) summarization is a critical yet cognitively demanding clinical documentation task that synthesizes voluminous electronic health records into concise discharge narratives. While recent large language models have demonstrated promise for automated clinical summarization, existing approaches exhibit significant limitations: single-pass generation lacks mechanisms for iterative refinement, leading to factual hallucinations; sequential architectures fail to model complex temporal and causal dependencies between clinical events; and opaque generation impedes interpretability essential for clinical adoption. We introduce a new method named Graph-of-Thoughts Hierarchical Clinical Summarization, or GOT-HCS. This method solves all these problems by formalizing the process and enhancing the knowledge. It comprises a directed graph of thoughts in which every node implies one of the specific steps of the clinical reasoning process. The node edges are time relationships, cause and logical relationships. Graph attention networks and knowledge graphs are also included in the approach. A multi-stage reasoning engine is developed by generating, aggregating and refining candidate summaries in an iterative way based on learned quality scoring, and hierarchical distillation structures thoughts into coherent narrative generative clinical themes. Extensive analysis of three clinical benchmarks MIMIC-IV-BHC, MTS-Dialog, MIMIC-III shows that GoT-HCS can expect to outperform GPT-4 by 6.5 % increase on hospital course summarization and 3.8 % decrease on MEDCON, as well as retain the competitive computational performance. The systematic ablations confirm the role of graph-based reasoning, knowledge buildup and refinement as significant aspects of performance and a qualitative examination indicates better preservation of information and fewer hallucinations than sequential generation methods. The source code is publicly available at https://github.com/AsmaAlhashmi01/GOT-HCS/