Large language models and knowledge graphs: solutions for health data management
May 2024. A proposal that combines the use of large language models with data graphs for clinical data: this is the innovation proposed in the article Augmented non-hallucinating large language models as medical information curators published in NPJ Digital Medicine, researchers Stephen Gilbert and Jakob Nikolas Katherfrom the Else Kröner Fresenius Center for Digital Health at Dresden University of Technology, Germany, together with the deputy director of the Millennium Institute Foundational Research on Data, Aidan Hogan, from the Department of Computer Science at the University of Chile.
Medical data presents a significant challenge for current data science: it is complex, must necessarily be protected, and there are a number of challenges to its efficient management. All this complexity has promoted the search for alternatives that would enable its application and make this valuable information can be used by healthcare providers for the benefit of their patients.
Large language models (LLMs) —very large deep learning models trained with large amounts of data, the basis for applications such as ChatGPT—can contribute significantly to better structuring, categorization, and interpretation of medical information. However, they have weaknesses that hinder their use in an area as critical as healthcare. The generation of plausible but incorrect information (hallucinations), or the ability to give different answers to the same query, make their use in medicine very complex. On the other hand, knowledge graphs are a way of structuring large amounts of information that may be in various formats or multivariable, through the use of nodes and vertices that connect the data.
Researchers at the EKFZ for Digital Health at TU Dresden and the University of Chile propose in the article Augmented non-hallucinating large language models as medical information curators a possible solution to this problem: the combination of LLM with knowledge graphs (KG). This gives rise to a new form of retrieval-augmented generation, Retrieval Augmented Generation, which would allow models to be more reliable, robust, and capable of reproducing queries.
The reliable recording of medical information and its exchange between different systems (interoperability) is a major challenge in healthcare and is often referred to as the "communication problem" in medicine. Medical ontologies and knowledge graphs (KGs) are approaches to solving this problem. Medical ontologies function as dictionaries of medical terms that help categorize and define medical concepts. However, since terms in human language can have different meanings depending on the context, these ontologies are often ambiguous. The word "cold," for example, can refer to body temperature, environmental conditions, or a cold. The same is true in all languages, and there are differences between different disciplines within healthcare. The use of acronyms is another major challenge in this field: COLD can also mean "chronic obstructive lung disease."
Knowledge graphs (KGs) are organized networks that connect different medical concepts and their relationships. For example, the term "COVID-19" in a graph could be connected to "fever" through a link called "has symptom."Graphs facilitate the understanding and processing of medical information, but they face challenges similar to those of medical ontologies.
Combination for structured reasoning
To remedy these shortcomings, researchers in Dresden and Santiago, Chile propose combining LLMs with KGs, leveraging their respective strengths. This combination provides structured reasoning and could help reduce model bias and deliver more reliable, accurate, and reproducible results. These approaches would be more compatible with regulatory approval pathways than LLMs alone.
"The combination of large language models and knowledge graphs is a way to link existing medical knowledge with the cognitive capabilities of large language models. We are only at the beginning of a very exciting development, "says Professor Jakob N. Kather, Chair of Clinical Artificial Intelligence at Dresden University of Technology and oncologist at Carl Gustav Carus University Hospital in Dresden.
In the research, the authors discuss different approaches to combining LLMs with KGs. They suggest that this could also facilitate the development of robust "digital twins" of patients, in the form of structured individual medical records that enable personalized diagnosis.
"Although regulatory challenges remain, healthcare professionals graduating today can anticipate access to compatible, advanced clinical information summarization tools that were unimaginable just five years ago. Furthermore, approaches that combine large language models with knowledge graphs are more likely to achieve early approval in conservative regulatory pathways," says Professor Stephen Gilbert, Chair of Regulatory Science for Medical Devices at the Technical University of Dresden.
For Aidan Hogan, "Like LLMs, KGs have applications not only in medicine, but also in many aspects of society that increasingly depend on data capture and processing. To make better decisions today, we must learn from the past, and the digital past is data. But integrating data comes at a higher cost, limiting this virtual perspective of the past. In this context, LLMs help integrate large-scale text data, while KGs help integrate large-scale structured data. Both approaches are complementary, and their combination can have many applications in different areas where data-driven decisions are made."
Interoperability of clinical data
The interoperability of clinical data is a major issue worldwide, and Chile is no exception. There have recently been advances in legislation in this regard, aimed at promoting the portability of patient data and medical records between different healthcare institutions: the idea is that when a patient visits any clinic, hospital, or healthcare center, it should be possible to share their medical history in digital format, accurately and completely, in compliance with applicable privacy and use regulations and ethics, without errors or omissions that could even be fatal. "Currently, LLMs have a great capacity to integrate text from different sources, but they do not have the necessary accuracy for this use case. So we propose that LLMs should be combined with other methods—in particular KGs—so that they can be used in the context of clinical data interoperability and other medical applications," Hogan points out.