RP17 - Application of Graph Neural Networks (GNN) for the Analysis of Patient Graphs in FHIR Format in Patients with Melanoma

The analysis of patient data from routine clinical practice plays an important role in medical research and the improvement of diagnostic and treatment methods. Recently, the Fast Healthcare Interoperability Resources (FHIR) format has gained a lot of importance in clinical medicine as it allows a standardized and interoperable representation of patient data. A patient dataset in FHIR format can be viewed as a directed graph, where nodes and edges represent different information and relationships.

Despite the inherent graphical nature of FHIR datasets, the graph structure has often been ignored in medical research. Instead, data is most often extracted from the FHIR graph in tabular form or as a timeline, largely discarding information about the structural makeup of the graph. This approach can result in potentially important relationships and patterns encoded in the graph structure going undetected. Therefore, it is of great importance to pay more attention to the structure of patient graphs in FHIR format and to develop innovative analysis methods to detect the relationships hidden within them. Graph neural networks (GNNs) are a promising method for analyzing FHIR data because they are specifically designed to process and learn from graph structures1. Unlike traditional neural networks, which are primarily designed to handle gridded or sequential data, GNNs can directly leverage the inherent graph structure of FHIR datasets to effectively model complex relationships and patterns within the data. Through their ability to process the node and edge information of a graph, GNNs enable a better representation of the patient information encoded in FHIR data and the relationships between different medical entities, such as diagnoses, treatments, or laboratory results. These graph-based models can help uncover previously undiscovered patterns and relationships that may be critical for improved patient stratification, personalized therapy approaches, and more accurate prediction of disease progression. In addition, GNNs are able to capture both local and global structural information within the graph, which can lead to more robust and informative analysis. This aspect makes GNNs particularly well-suited for FHIR data analysis, as they can not only take advantage of graph representation, but also account for the complexity and heterogeneity of data often encountered in medical research and patient care.

In the proposed research project, we are investigating the use of GNNs to analyze patient graphs in FHIR format in patients with melanoma. The goal is to facilitate the identification of similar patient profiles and the creation of synthetic cohorts, as well as to impute missing data in FHIR format.

Integrating GNNs to analyze FHIR data in WisPerMed will help uncover previously untapped information and relationships in the complex and heterogeneous medical data. This will make it possible to better identify and compile contextually relevant facts for treatment decisions. By combining methods from information extraction, knowledge representation with machine learning techniques and insights into user interaction at the point of care (PoC), the project will contribute to the development of new, interdisciplinary approaches in medical informatics.

[1] L. Wu, P. Cui, J. Pei, and L. Zhao. Graph Neural Networks: Foundations, Frontiers, and Applications. Springer, Singapore, 2022

[2] C. Yang, P. Zhuang, W. Shi, A. Luu, L. Pan. Conditional structure generation through graph variational generative adversarial nets. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2019