Clinical Language Models for Information Extraction and Predictive Tasks from Clinical Notes: Uncovering the Potential of Unstructured Clinical Data

• Autor(a) principal: Lopes, Ana Filipa Gonçalves
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/163653
• topic: Natural Language Processing; Machine Learning; Clinical Language Models; Electronic Health Record Clinical Notes; ICD-9 code extraction; Cardiothoracic Surgery; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

The exponential growth of electronic health records has resulted in an unprecedented volume of unstructured clinical data. Harnessing the potential of this information requires advanced natural language processing techniques and holds immense potential for healthcare improvement. This dissertation explores the potentialities of clinical language models in extracting and organizing the information from clinical text, and aims to measure the impact of such information in a real clinical task involving the prediction of complications following Cardiothoracic Surgery. Three information extraction models were developed by fine-tuning clinical language models on a ICD-9 code classification task. The ClinicalBERT and BioGPT-based models achieved a Mean Average Precision at 10 around 0.437, outperforming a fine-tuned model from the literature. These models were later applied to extract ICD-9 codes from translated Portuguese clinical notes. The retrieved variables were proven to benefit machine learning models trained to predict post-surgery complications, as their accuracy improved in up to 30% relative to a baseline model not trained with this information, when the information extracted from clinical text was added, achieving values of around 0.880. The results of this research solidify clinical language models as powerful tools for clinically relevant information extraction from free-text medical reports, and set the tone for the integration of these systems in clinical decision support systems, towards high performance and interpretability standards.