Caracterización de un corpus extraído de historias clínicas electrónicas de maternas a través de técnicas de procesamiento de lenguaje natural / Characterization of a corpus extracted from maternal electronic health records through natural language processing techniques

Este artículo tuvo como propósito caracterizar el texto libre disponible en una historia clínica electrónica de una institución orientada a la atención de pacientes en embarazo. La historia clínica electrónica, más que ser un repositorio de datos, se ha convertido en un sistema de soporte a la toma de decisiones clínicas. Sin embargo, debido al alto volumen de información y a que parte de la información clave de las historias clínicas electrónicas está en forma de texto libre, utilizar todo el potencial que ofrece la información de la historia clínica electrónica para mejorar la toma de decisiones clínicas requiere el apoyo de métodos de minería de texto y procesamiento de lenguaje natural. Particularmente, en el área de Ginecología y Obstetricia, la implementación de métodos del procesamiento de lenguaje natural podría ayudar a agilizar la identificación de factores asociados al riesgo materno. A pesar de esto, en la literatura no se registran trabajos que integren técnicas de procesamiento de lenguaje natural en las historias clínicas electrónicas asociadas al seguimiento materno en idioma español. En este trabajo se obtuvieron 659 789 tokens mediante los métodos de minería de texto, un diccionario con palabras únicas dado por 7 334 tokens y se estudiaron los n-grams más frecuentes. Se generó una caracterización con una arquitectura de red neuronal CBOW (continuos bag of words) para la incrustación de palabras. Utilizando algoritmos de clustering se obtuvo evidencia que indica que palabras cercanas en el espacio de incrustación de 300 dimensiones pueden llegar a representar asociaciones referentes a tipos de pacientes, o agrupar palabras similares, incluyendo palabras escritas con errores ortográficos. El corpus generado y los resultados encontrados sientan las bases para trabajos futuros en la detección de entidades (síntomas, signos, diagnósticos, tratamientos), la corrección de errores ortográficos y las relaciones semánticas entre palabras para generar resúmenes de historias clínicas o asistir el seguimiento de las maternas mediante la revisión automatizada de la historia clínica electrónica(AU)

The purpose of this article was to characterize the free text available in an electronic health record of an institution, directed at the care of patients in pregnancy. More than being a data repository, the electronic health record (HCE) has become a clinical decision support system (CDSS). However, due to the high volume of information, as some of the key information in EHR is in free text form, using the full potential that EHR information offers to improve clinical decision-making requires the support of methods of text mining and natural language processing (PLN). Particularly in the area of gynecology and obstetrics, the implementation of PLN methods could help speed up the identification of factors associated with maternal risk. Despite this, in the literature there are no papers that integrate PLN techniques in EHR associated with maternal follow-up in Spanish. Taking into account this knowledge gap, in this work a corpus was generated and characterized from the EHRs of a gynecology and obstetrics service characterized by treating high-risk maternal patients. PLN and text mining methods were implemented on the data, obtaining 659 789 tokens and a dictionary with unique words given by 7 334 tokens. The characterization of the data was developed from the identification of the most frequent words and n-grams and a vector representation of embedding words in a 300-dimensional space was performed using a CBOW (Continuous Bag of Words) neural network architecture. The embedding of words allowed to verify by means of Clustering algorithms, that the words associated to the same group can come to represent associations referring to types of patients, or group similar words, including words written with spelling errors. The corpus generated and the results found lay the foundations for future work in the detection of entities (symptoms, signs, diagnoses, treatments), correction of spelling errors and semantic relationships between words to generate summaries of medical records or assist the follow-up of mothers through the automated review of the electronic health record(AU)

Improving the CONTES method for normalizing biomedical text entities with concepts from an ontology with (almost) no training data

Entity normalization, or entity linking in the general domain, is an information extraction task that aims to annotate/bind multiple words/expressions in raw text with semantic references, such as concepts of an ontology. An ontology consists minimally of a formally organized vocabulary or hierarchy of terms, which captures knowledge of a domain. Presently, machine-learning methods, often coupled with distributional representations, achieve good performance. However, these require large training datasets, which are not always available, especially for tasks in specialized domains. CONTES (CONcept-TErm System) is a supervised method that addresses entity normalization with ontology concepts using small training datasets. CONTES has some limitations, such as it does not scale well with very large ontologies, it tends to overgeneralize predictions, and it lacks valid representations for the out-of-vocabulary words. Here, we propose to assess different methods to reduce the dimensionality in the representation of the ontology. We also propose to calibrate parameters in order to make the predictions more accurate, and to address the problem of out-of-vocabulary words, with a specific method.