Automatic text processing of clinical narratives

• Autor(a) principal: Faustino, Jorge Filipe Balseiro
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/38432
• topic: Concept recognition; Information retrieval; Natural language processing; Text mining

The informatization of medical systems and the subsequent move towards the usage of Electronic Health Records (EHR) over the paper format by medical professionals allowed for safer and more e cient healthcare. Additionally, EHR can also be used as a data source for observational studies around the world. However, it is estimated that 70-80% of all clinical data is in the form of unstructured free text and regarding the data that is structured, not all of it follows the same standards, making it di cult to use on the mentioned observational studies. This dissertation aims to tackle those two adversities using natural language processing for the task of extracting concepts from free text and, afterwards, use a common data model to harmonize the data. The developed system employs an annotator, namely cTAKES, to extract the concepts from free text. The extracted concepts are then normalized using text preprocessing, word embeddings, MetaMap and UMLS Metathesaurus lookup. Finally, the normalized concepts are converted to the OMOP Common Data Model and stored in a database. In order to test the developed system, the i2b2 2010 data set was used. The di erent components of the system were tested and evaluated separately, with the concept extraction component achieving a precision, recall and F-score of 77.12%, 70.29% and 73.55%, respectively. The normalization component was evaluated by completing the N2C2 2019 challenge track 3, where it achieved a 77.5% accuracy. Finally, during the OMOP CDM conversion component, it was observed that 7.92% of the concepts were lost during the process. In conclusion, even though the developed system still has margin for improvements, it proves to be a viable method of automatically processing clinical narratives.