Desarrollo de una aplicación web que permita identificar automáticamente los principales tópicos en documentos médicos digitales

Abstract

Abstract: Topic Modeling is an unsupervised machine learning technique that allows to extract topics from certain documents and classify them. In this work, a web application was developed based on the Latent Dirichlet Allocation (LDA) model to extract topics from documents with medical information. 220 digital documents were used, these documents have information about health conditions of different patients. A pre-processing was carried out, which implies cleaning and refining the data to define the medical data or terms that can represent the documents. Subsequently, a document representation was made through a document term matrix. LDA was applied with this matrix, therefore the document-topic matrix and topic-term matrix were obtained. A very important feature that was added to the application is to give options to choose the number of the topics. For this, 4 metrics based on graphs were used to allow the selection of the number of topics. Also, the use of specialized medical terminology turned out to be a better option when selecting the type of vocabulary that describes the documents. The results can be evidenced by a heat map, word cloud and an interactive tool called PyLDAvis.