Ontology-Constrained Generation of Domain-Specific Clinical Summaries

TL;DR

This study proposes a novel approach that leverages ontologies to create domain-adapted summaries both structured and unstructured for clinical notes and hallucination reduction, employing an ontology-guided constrained decoding process to reduce hallucinations while improving relevance.

Abstract

Large Language Models (LLMs) offer promising solutions for text summarization. However, some domains require specific information to be available in the summaries. Generating these domain-adapted summaries is still an open challenge. Similarly, hallucinations in generated content is a major drawback of current approaches, preventing their deployment. This study proposes a novel approach that leverages ontologies to create domain-adapted summaries both structured and unstructured. We employ an ontology-guided constrained decoding process to reduce hallucinations while improving relevance. When applied to the medical domain, our method shows potential in summarizing Electronic Health Records (EHRs) across different specialties, allowing doctors to focus on the most relevant information to their domain. Evaluation on the MIMIC-III dataset demonstrates improvements in generating domain-adapted summaries of clinical notes and hallucination reduction.

Figures (5)

• Figure 1: Overview of our general architecture to generate domain-tailored summaries.
• Figure 2: Overall architecture of our method : Multiple notes about the same patient are passed to the framework and structured and unstructured summaries are generated
• Figure 3: Extraction Phase
• Figure 4: Constrained decoding process : Each beam rectangle represents the current generation window associated with the beam. The concepts in green are concepts that are associated to a children class of the base class (Drug or medicament) in the ontology. Green concepts improve the hierarchy score which augments the probability that the beam will be chosen as a final output. The similarity score is computed using the ROUGE-2 score between the generation window and the clinical notes.
• Figure 5: Most frequent classes in SNOMED-CT based on different domains of clinical notes. This plot was computed by automatically annotating 1000 clinical notes from each category in MIMIC-III and associating each concept to a SNOMED-CT class. We then performed a domain adaptation analysis as shown in section \ref{['sec:domain_analysis']}. Only the 5 most frequent concepts were kept in this figure.