A semantic interoperability model based on NLP for nonstructured health data

Abstract

The healthcare domain faces significant challenges in managing the rapidly growing volume of data generated daily, particularly in the collection and sharing of this information. Healthcare professionals such as physicians, nurses, radiologists, cardiologists, surgeons, and other specialists frequently enter patient data into electronic systems, often in an open, unstructured textual format. We conducted a literature review that reveals several challenges in processing real-world data, with one critical issue being the scarcity of tools and dictionaries available in Portuguese for the healthcare sector. This gap, coupled with the unique challenges inherent in healthcare data processing, adds considerable complexity to extracting and structuring essential information from clinical records. Additionally, ensuring data interoperability between different healthcare providers becomes challenging when these providers do not initially aim for interoperability during input data. Observing these challenges, this research proposed a model to enable semantic interoperability of clinical notes from electronic health record systems. The methodology used in this research has an applied and exploratory character, and it has been evaluated through the development of a prototype. This approach aims to address some of the current limitations in data processing and integration, specifically within the Portuguese healthcare context, and to create a flexible model that can treat real-world data more effectively in structuring and sharing data. This research is part of the MyDigitalHealth project, a collaboration between the university and six hospitals in Porto Alegre, which provided data from hospitalized patients who tested positive for COVID-19, ensuring a real-world context for data issues. We analyzed the characteristics of the data with respect to interoperability between providers and proposed a model that involves hybrid techniques for information extraction, lexical normalization, and structure for standard harmonization. Thus, we defined a set of experiments using machine learning, combining the Transformers architecture for entity recognition with natural language processing for lexical normalization and semantic matching and adopting OWL ontologies as an intermediary representation structure. The experiments revealed three main contributions. First, we developed a specialized annotated dataset, classifying six entities with 18,666 validated annotations by specialists in 314 documents. Second, we conducted experiments using BERT models fine-tuned on our small dataset for entity recognition, achieving 95% accuracy, with precision rates of 90% for classifying entities related to Invasive or Therapeutic Procedures and 89% for Disease or Syndrome and Diagnostic Procedures. These results demonstrate the model’s effectiveness in extracting relevant information from unstructured clinical notes. Third, ontologies as intermediary representation structures ensured semantic consistency and enhanced interoperability in an independent format. The limitations and opportunities for future studies from this research include applying the model to data from different domains, such as nursing notes, odontology, clinic context, and accountability records. Another topic is the gap in term disambiguation and semantic alignment in healthcare data, focusing on linking terminologies to structured data, ensuring international coding for clinical data, and enabling interoperability across borders. Finally, this research aims to contribute to the continuity of citizen healthcare and guide developers and providers in building robust and complex platforms that implement the use of healthcare standards. We also expect more and more professionals and health managers to improve healthcare worldwide through the adoption of international standards within electronic health record systems.