SPEER (Sentence-level Planning via Embedded Entity Retrieval) is a method proposed to aid clinicians in writing lengthy hospital discharge summaries. The task of summarizing a patient’s hospital stay is time-consuming due to the number of unique clinical concepts involved. Current open-source LLMs (Large Language Models) were found to provide incomplete and unfaithful summaries. To improve this, SPEER was developed, which uses sentence-level planning to ensure salient entities are covered, improving the coverage and faithfulness of these summaries. The method was fine-tuned on a large-scale dataset of ~167k in-patient hospital admissions and showed gains over non-guided and guided baselines.
Publication date: 4 Jan 2024
Project Page: https://arxiv.org/abs/2401.02369
Paper: https://arxiv.org/pdf/2401.02369
