Um modelo multinível para estruturação de informações contidas em evoluções de prontuários escritos em texto livre

Abstract

The medical field has undergone a series of transformations with the adoption of new technologies. One aspect that has seen significant changes is how patient information is stored. Electronic health records have brought a series of advantages, but they still present some issues. One of these issues is of structuring of the information contained in clinical notes. These notes can be stored in free text, that is, in an unstructured form; in a semi-structured form, containing a list of free-text fields to categorize each piece of information; in a structured form, where each piece of information has a series of specific fields; or a combination of these forms. Greater structuring brings a wealth of information and ease of automated consultation. On the other hand, the medical staff must dedicate more attention when managing clinical notes due to the rules to maintain its structure. This problem is more evident when migrating from a less structured record to a more structured record, given the impracticality of direct migration. Considering these aspects, this study arises from a concrete need, related to the migration of an electronic health record software from a company that used unstructured clinical notes to an electronic health record with structured and semi-structured records. For this migration to be effective, the following software requirements are imperative: that all relevant information is maintained and that it is at least semi-structured and, when possible, fully structured. Through a systematic review to find the state of the art in the field, no proposal was found that satisfactorily meets these requirements. Considering this context, this work proposes a multi-level model for structuring progress notes written in free text in the Portuguese language. The main requirements of the model are that, in the structuring process, all relevant information from the clinical notes is maintained, that the information is structured at least at the sentence level, and that, when possible, each entity in the sentence is also structured. The model consists of a pipeline with two main components: the first is responsible for structuring the information at the sentence level, by dividing the text and individually classifying each sentence in the SOAP notes. At the second level, if possible, structuring is done at the level of its entities, identifying and relating them. To evaluate the viability of the model, a prototype of the pipeline was implemented, using natural language processing and machine learning techniques, such as BERT models, associating its subcomponents with classic NLP tasks such as sentence boundary detection, sentence classification, named entity recognition, relation extraction, and ontology matching. For training and evaluating the pipeline and networks, a database provided by the company that motivated the study was used. The database contains 10,000 records and 234,673 clinical notes, of which 15,883 were divided into 100,021 sentences, classified, and structured through a Graphical User Interface (GUI) developed for this task, forming the gold standard for network training. After a series of training and evaluations, the best networks were selected, and the pipeline was implemented. For the final evaluation of this pipeline, 721 records with a total of 10,013 sentences were used, which were also classified using the GUI, forming the gold standard for the final evaluation. The results obtained were compared with those of the pipeline, achieving an accuracy of 0.8641, precision of 0.9493, and F-score of 0.9029 for the first level of structuring, and an accuracy of 0.8354, precision of 0.9382, and F-score of 0.8815 for the second level of structuring.