Best Paper Selection

Appendix: Content Summaries of Selected Best Papers for the 2024 IMIA Yearbook, Section Knowledge Representation and Management

Ricardo M. S. Carvalho, Daniela Oliveira, Catia Pesquita

Knowledge Graph Embeddings for ICU readmission prediction

BMC Med Inform Decis Mak 2023;23:12

In this paper, the authors developed an approach to predicting Intensive Care Unit (ICU) readmissions by integrating semantic annotations as Knowledge Graphs (KGs) into Electronic Health Record (EHR) data. The authors take the opportunity to link EHR data to ontologies. They can thus take into account the data context and feed this extra later of information about the meaning of the data to machine learning systems. The authors then apply more conventional learning algorithms to test whether their KG approach classifies ICU patient readmission risks better than other methods.The interest of this article is to develop a complete process starting from a MIMIC-III patient database, annotating it with ontologies/terminologies (NCIt, SNOMED, MeSH, ICD-9, RxNorm, etc. thus producing KGs representing patients. These KGs are then vectorized using embedding techniques in a way that machine learning algorithms can process. It should be noted that the quality of semantic annotation is not directly tested, but the paper is very comprehensive in verifying the overall performance of their system as a function of the different ontologies used and the different embedding algorithms. As written by the authors, this work demonstrates the potential for impact that integrating ontologies and KGs into biomedical machine learning applications can have. Moreover, by having the clinical data semantically annotated with ontologies, this work also paves the way for more explainable approaches that explore the meaning encoded in ontologies to better explain predictions to clinicians. It remains to be seen how such processes would be set up in a hospital and for other patient data processing tasks.

J Harry Caufield, Tim Putman, Kevin Schaper, Deepak R Unni, Harshad Hegde, Tiffany J Callahan, Luca Cappelletti, Sierra A T Moxon, Vida Ravanmehr, Seth Carbon, Lauren E Chan, Katherina Cortes, Kent A Shefchek, Glass Elsarboukh, Jim Balhoff, Tommaso Fontana, Nicolas Matentzoglu, Richard M Bruskiewich, Anne E Thessen, Nomi L Harris, Monica C Munoz-Torres, Melissa A Haendel, Peter N Robinson, Marcin P Joachimiak, Christopher J Mungall, Justin T Reese.

KG-Hub-building and exchanging biological knowledge graphs

Bioinformatics 2023;39(7):btad418

The authors highlight the value of KGs for integrating heterogeneous data from different sources and in various domains including biology, ecology, biomedicine, and personalized health. But they note that, despite the demonstrated usefulness of KGs, barriers exist (mainly non-standard formats) that limit their effectiveness and reusability. They present a solution, KG-Hub, a platform that accelerates the construction and reuse of KGs by developing tooling and design patterns to encourage FAIRness. KG-Hub proposes thus a wide range of services as the possibility to incorporate ontologies in KG, graph machine learning on KGs, harmonizing data sources (with Biolink as pivot model), and data downloading, transformation, and graph assembly (with KGX (Knowledge Graph Exchange) as the KG serialization standard). For requesting and interfaces, Neo4J is in the limelight because it is flexible enough to interface with a wide range of formats. Finally, the authors have implemented an extremely comprehensive tutorial for setting up a KG-Hub project, loading KGs and using them for various tasks. These tutorials provide practice examples of how KG-Hub can be used, especially for downstream graph machine learning use cases such as node embedding and link prediction using GRAPE, and automated graph machine learning using NEAT. It's worth noting that all the source codes for the projects displayed are available.

Mingzhou Fu, Yu Yan, Loes M Olde Loohuis, Timothy S Chang

Defining the distance between diseases using SNOMED CT embeddings

J Biomed Inform 2023;139:104307

Previous methods for measuring disease distance only worked for small sets of diseases. This research addresses that gap by creating a method that works for all diseases listed in ICD-10 (International Classification of Diseases, version 10). The proposed distance metric is based on the KG embeddings of SNOMED CT. It was compared to three other metrics: ICD tree-based distance, clinical comorbidity-based distance, and genetic correlation-based distance. The SNOMED CT embedding-based metric used the RotatE model to obtain embedding representations of SNOMED CT concepts and calculated pairwise similarities between three-digit ICD-10 codes. The ICD tree-based metric utilized the hierarchical structure of ICD-10 codes to determine distances based on chapter and numeric differences. The clinical comorbidity-based metric used patient health records to measure dissimilarity between diseases based on their co-occurrence. The genetic correlation-based metric estimated genetic correlations using linkage disequilibrium score regression (LDSC) based on GWAS summary statistics from the UK Biobank and mined correlations reported in published literature. The results showed that the SNOMED CT embedding-based metric aligned well with the hierarchical structure of ICD codes and effectively captured more fine-grained disease relationships. Furthermore, even without direct involvement of genetic or clinical data for the metric construction, the SNOMED CT embedding-based metric could still capture significant genetic and clinical co-occurrence information. Overall, this study provides a comprehensive analysis of disease distance metrics and highlights the advantages of the embedding-based metric in capturing disease relationships at a finer granularity, providing valuable insights for disease characterization and personalized medicine.

Die Autoren geben an, dass kein Interessenkonflikt besteht.

Publikationsverlauf

Artikel online veröffentlicht:
08. April 2025