Healthcare data interoperability with Large Language Models

Data interoperability is crucial for making the most of available information, and it becomes even more valuable when it comes to healthcare data. In this field, collaboration and interaction among different medical environments are key to quickly and effectively managing diseases. Moreover, the collaboration of various medical organizations and the smooth exchange of information among them can be used to improve research quality and clinical practice. However, this task is quite complex due to the great diversity of the data collected and its different nature. This information involves both structured and unstructured data, capturing all clinical events of patients. Medical interoperability standards emerge to try to mitigate and solve this problem by representing clinical information in a standardized manner and common format. However, there is a wide range of interoperability standards in this field, which creates significant difficulty when exchanging information. Standards such as HL7 FHIR [1], OMOP, HL7 CDA, etc., lead to a lack of a clear standardization goal since each organization can choose any of the existing standards. Additionally, the process of standardizing medical data is very challenging and complex. This is because those responsible for this transformation must have deep knowledge of the medical field, as well as extensive experience and understanding of the interoperability standard to be used. For these reasons, and due to the emergence of disruptive technologies like Large Language Models (LLMs), it is proposed to leverage these tools to address this issue and try to provide a method that eases or even overcomes this challenge. In this context, this work is being developed. A methodology is being created that is capable of representing structured medical data using a medical standard, specifically, HL7 FHIR, through the use of LLMs. The methodology has been evaluated using the MIMIC-IV dataset [2], starting from a single table where each column represents an aspect of the data. Its goal is as follows: starting from a single table that compiles clinical information, where each row represents an attribute of this clinical data, the methodology transforms the data into HL7 FHIR. In this way, the corresponding HL7 FHIR model has been identified for each row; that is, which resource and attributes are needed to represent the input clinical information. To achieve this, different techniques such as unsupervised clustering and prompt engineering combined with Retrieval Augmented Generation (RAG) are implemented to interact with the LLM. The results obtained in this process show a 94% accuracy in identifying the correct HL7 FHIR resources and a 70.15% accuracy in identifying the right HL7 FHIR attributes. These are very promising results for this challenge. Currently, this methodology is being applied in a real use case to study its impact. All of this is part of the European project (https://www.resqplus.eu), which aims to standardize a stroke registry using HL7 FHIR and SNOMED CT, thereby facilitating its interoperability on an international level. However, there are several limitations and challenges that are proposed to be addressed in order to increase both the robustness and quality of the methodology. The main issue identified is the lack of a broad context that explains the source data, which hinders the performance and generalization of the LLM. It is also necessary to mention the difficulty in relating the different resources, since each HL7 FHIR resource has many interrelationships. This creates a need for a global view of the data before carrying out the process of identifying resources and attributes related to the interoperability standard.Regarding the future of this work, improvement paths are proposed, such as implementing support for a larger number of medical interoperability standards (HL7 CDA, OMOP, etc.) and allowing the use of unstructured data, which would considerably increase the tool’s adaptability. Additionally, finetuning processes for the LLM are being considered, which would give the model a greater capacity for generalization and ensure improved performance. Moreover, the use of opensource LLMs is proposed to enable the platform to scale and the tool to be more customizable. The development of a user-friendly and intuitive interface is currently underway with the aim of making a tool accessible to any type of user. With the advances described and the future directions proposed, this work proposes and implements a method for automating or semi-automating the process of medical data interoperability. Therefore, it increases the possibility of conducting studies and research using data from various medical sources, as well as improves the interaction among different health entities—ultimately enhancing the quality of treatment and patient care.