RDBU| Repositório Digital da Biblioteca da Unisinos

Anamtech: speech-based automatic structuring of medical anamnesis

Show simple item record

metadataTrad.dc.contributor.author Fraga, Ygor Allan de;
metadataTrad.dc.contributor.advisor Costa, Cristiano André da;
metadataTrad.dc.contributor.advisorLattes http://lattes.cnpq.br/9637121030877187;
metadataTrad.dc.publisher Universidade do Vale do Rio dos Sinos;
metadataTrad.dc.title Anamtech: speech-based automatic structuring of medical anamnesis;
metadataTrad.dc.description.resumo The medical history process is critical for a correct diagnosis of the patient. Filling out medical documents is costly for the doctor and can cause some conversation details to go unnoticed, resulting in a bad patient experience or a wrong diagnosis. Helping the experience of both physician and patient is the motivation behind this work. The main objective is to create an application that automatically integrates speech recognition to turn into text the interview, identify the relevant entities for the anamnesis document, and structure a digital document. The developed model (a.k.a. Anamtech) integrates different services to make it possible to recognize the anamnesis properly automatically. Voice recognition was used to capture the conversation between doctor and patient. Several open libraries have already transposed the audio into text. The recognized text was included in the process of identifying essential terms for anamnesis, which healthcare professionals reviewed, and an entity recognition algorithm was used to identify such information. This algorithm was previously trained according to available existing anamnesis that passed through the process of labeling. The last Anamtech component organizes all the recognized entities in a document following a defined medical standard. A complete automatic application was created and ready to use with minor interference by the physician who uses it. As the final document is divided by entities, organized with a prefix by the anamnesis phase, it would be easy to change any information contained in it. In general, the named entity recognition (a.k.a. NER) model, which is the heart of this project, had a precision of 85.1%, a recall of 87.6%, and an f1-score of 86.3%. In addition, metrics for each one of the entities were captured and described. The metrics related to the patient identification had the best results, whereas the ones associated with symptoms, diseases, and treatments could be identified, but some mismatches were identified due to the difficulty to classify some entities in the pre-processing.;
metadataTrad.dc.subject Speech recognition; Named entity recognition; Anamnesis; Natural Language Processing; Medical Informatics;
metadataTrad.dc.type TCC;
metadataTrad.dc.date.issued 2021-12-09;
metadataTrad.dc.identifier.uri http://repositorio.jesuita.org.br/handle/UNISINOS/13270;
metadataTrad.dc.audience.educationLevel Graduação;
metadataTrad.dc.curso Sistemas de Informação;


Files in this item

This item appears in the following Collection(s)

Show simple item record

Search

Advanced Search

Browse

My Account

Statistics