MultiSynDS: large scale multilingual synthetic discharge summary generation by means of generative LLMs and clinical case report transformation

Access to real clinical records is often restricted because of privacy and legal constraints, particularly for languages other than English. This limitation reduces the availability of structured data needed to develop and evaluate multilingual clinical NLP systems. This thesis introduces MultiSynDS...

Descripción completa

Detalles Bibliográficos
Autor: Becerra Tomé, Alberto
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/446598
Acceso en línea:https://hdl.handle.net/2117/446598
Access Level:acceso abierto
Palabra clave:Medical records
Natural language processing (Computer science)
Datos sintéticos
PLN clínico
PLN multilingüe
Informes de alta
Modelos de lenguaje a gran escala
Evaluación de modelos
Generación de lenguaje natural
LLM como juez
Synthetic data
Clinical NLP
Multilingual NLP
Discharge summaries
Large language models
Model evaluation
Natural Language Generation
LLM as a Judge
Històries clíniques
Tractament del llenguatge natural (Informàtica)
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Llenguatge natural
Descripción
Sumario:Access to real clinical records is often restricted because of privacy and legal constraints, particularly for languages other than English. This limitation reduces the availability of structured data needed to develop and evaluate multilingual clinical NLP systems. This thesis introduces MultiSynDS, a pipeline that generates synthetic, multilingual discharge summaries from publicly available clinical case reports. Leveraging generative large language models, the approach transforms free-text narratives into structured summaries and produces translations in multiple languages. The investigation is organized around five research questions: first, whether a generative model can produce clinically acceptable discharge summaries; second, which factors complicate human evaluation of these summaries; third, how well concept-level automatic metrics align with expert judgments; fourth, the utility of the LLM-as-a-judge method and the influence of model size and domain adaptation; and fifth, whether models adapted to the medical domain improve agreement with expert evaluations compared to general-purpose models. The method was applied to 1,000 cardiology case reports from the AI4HF project. Synthetic summaries were assessed via clinical named-entity recognition, SNOMED CT concept linking, expert reviews on a 1¿5 scale, and rubric-based scoring by LLMs. For 30 selected summary pairs, human annotators assigned mean ratings of 4.09 for medical entities completeness, 4.49 for structure-header, 4.02 for structure-content, 4.21 for content accuracy, 4.25 for made-up content detection, and an overall quality score of 3.63. The findings suggest that concept-based automatic metrics correlate well with expert evaluations on core content dimensions, whereas both experts and LLMs exhibit lower agreement on more subjective aspects such as hallucinations and overall score. Although this work focuses on English, the pipeline is inherently multilingual and can be extended to other languages included in the DataTools4Heart project-namely Spanish, Dutch, Italian, Romanian, Czech, and Swedish. The full MultiSynDS dataset, comprising original clinical cases and generated discharge summaries in English, Spanish, and Dutch, is publicly available at https://zenodo.org/records/15664866. The codebase is also shared with the open-source community on GitHub at https://github.com/nlp4bia-bsc/MultiSynDS.