Applications of LLMs and Knowledge Graphs in Healthcare

AAAI-25 Bridge Program 25-26 February, 2025 Philadelphia, Pennsylvania, USA Applications of LLMs and Knowledge Graphs in Healthcare Enayat Rajabi 1,2 1 Cape Breton University, NS, Canada 2 Center for Applied Intelligent Systems Research (CAISR), Halmstad University, Sweden 1 AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 Presentor ●Associate Professor of Business Analytics, Cape Breton University, Canada ●Adjunct Professor, Faculty of Computer Science, Dalhousie University, Canada ●Prior Postdoctoral Fellow of Dalhousie University, Canada ●Ph.D. in Knowledge Engineering, University of Alcala, Spain ●Research interest in Knowledge Graphs and Machine Learning ○Knowledge-driven AI application in various domains including Healthcare 2 AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 Introduction ●Healthcare faces an explosion of scientific data: clinical trials, patient records, research papers. ●Traditional methods struggle to keep pace with this volume and complexity. 3 AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 Research Question “How to leverage Large Language Models (LLM) and Knowledge Graphs (KG) in Healthcare?” 4 AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 LLM+KG in Healthcare ●LLMs: ○Automate information extraction from vast text ○Summarize complex medical concepts ○Accelerate research by identifying key findings and patterns ●Knowledge Graphs: ○Reveal hidden connections and patterns within medical data. ○Enhance explainability, allowing clinicians to understand the reasoning behind AI-driven decisions. ○Enable question answering like: "What factors influenced the LLM's decision?" 5 Rajabi E, Etminani K. Knowledge-graph-based explainable AI: A systematic review. J Inf Sci. 2024 Aug;50(4):1019-1029. doi: 10.1177/01655515221112844. Epub 2022 Sep 24. PMID: 39135903; PMCID: PMC11316662. AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 Applications of LLM+KG in Healthcare ●Extracting Adverse Drug Events from Electronic Health Records ○Electronic Health Records (EHRs) contain a wealth of information about patient health, which is often buried in unstructured text, making it difficult to analyze and utilize for research or clinical decision-making. ○Example: a deep learning framework that combines KGs and pre-trained LLMs to predict drug-drug interactions (DDIs)1. 6 1Xu C, Bulusu KC, Pan H, Elemento O. DDI-GPT: Explainable Prediction of Drug-Drug Interactions using Large Language Models enhanced with Knowledge Graphs. bioRxiv [Preprint]. 2024 Dec 9:2024.12.06.627266. doi: 10.1101/2024.12.06.627266. PMID: 39713430; PMCID: PMC11661079. AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 LLM+KG in Drug-Drug Interaction 7 1Xu C, Bulusu KC, Pan H, Elemento O. DDI-GPT: Explainable Prediction of Drug-Drug Interactions using Large Language Models enhanced with Knowledge Graphs. bioRxiv [Preprint]. 2024 Dec 9:2024.12.06.627266. doi: 10.1101/2024.12.06.627266. PMID: 39713430; PMCID: PMC11661079. Using BioKG for sentence enrichment AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 LLM+KG for DDI: Example ● Doctor’s Note: "Patient was prescribed ibuprofen for pain relief. After three days, they developed gastric bleeding, requiring discontinuation of the medication." ● Using Named Entity Recognition (NER), the LLM identifies: ● Drug: Ibuprofen ● Adverse Event: Gastric bleeding ● Temporal Relation: "After three days" (suggesting causality) ● The KG contains known relationships: ● Ibuprofen → Can Cause → Gastrointestinal Bleeding (from medical databases like DrugBank, SIDER, or FDA reports) ● The KG cross-references patient history, checking for risk factors like pre-existing ulcers. ● The system generates an alert: "Potential Adverse Drug Event detected: Ibuprofen is associated with gastric bleeding in this patient. Consider alternative pain management." 8 AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 Classifying Medical Misinformation 9 Alber, D.A., Yang, Z., Alyakin, A. et al. Medical large language models are vulnerable to data-poisoning attacks. Nat Med (2025). https://doi.org/10.1038/s41591-024-03445-1 ● Example: Using biomedical knowledge graphs to screen medical LLM outputs Using KG to detect the misinformation AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 LLM Extraction Methods 16 While NER is a powerful method for extracting drugs, symptoms, and adverse events, other LLM-based approaches can further improve ADE detection: 1. Relation Extraction (RE) – Identifies cause-effect relationships between drugs and adverse effects. 2. Zero-Shot/Few-Shot Learning – Enables ADE detection without extensive labeled data. 3. Prompt Engineering – Uses structured prompts to extract and summarize ADEs. 4. Retrieval-Augmented Generation (RAG) – Integrates external medical databases (e.g., FAERS, DrugBank) with LLM responses. 5. Fine-Tuned Models – Tailored LLMs for clinical text analysis (e.g., BioBERT, MedGPT). AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 How LLM+KG help healthcare? 17 ● Retrieval-Augmented Generation (RAG): diagnostic predictions and clinical decision support, LLM+KG provide accurate and contextually relevant information. ● Knowledge Graph Construction and Completion: LLMs assist in building and enriching KGs by extracting entities and relationships from unstructured medical texts. ● Question Answering (QA) Systems: Combining LLMs with KGs enhances medical QA systems by providing precise and contextually relevant answers. ● Information Extraction and Entity Linking: LLMs, integrated with KGs, improve the extraction of medical information by accurately identifying and linking entities within medical texts. AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 Current Research 18 ● Following a LLM+KG approach to identify adverse drug events based on personal information of patients: ○ Gathering the source data (disease, drugs, protein, patients, genes, etc.) ○ Extract the important entities and relationships using relation extraction method ○ Constructing a KG ○ Predicting the adverse drug reaction events based on patient information (i.e. demographic data) AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 PrimeKG: A Precision Medicine-oriented KG 19 Chandak, P., Huang, K., & Zitnik, M. (2023). Building a knowledge graph to enable precision medicine. Scientific Data, 10(1), 67. 20 high-quality resources 17,080 diseases 4M relationships drug-disease prediction enable multi-modal analyses based on text descriptions of clinical guidelines and KG structure. Thank you Questions? [email protected] 20 AAAI Bridge on Artificial Intelligence for Scholarly Communication - February 2025 References 1. L. Ehrlinger and W. W oß. Towards a definition of knowledge graphs. SEMANTiCS (Posters, Demos, SuCCESS), 48(1-4):2, 2016. 2. G. Lagunes-Garcia, A. Rodriguez-Gonzalez, L. Prieto-Santamaria, E. P. Garcia del Valle, M. Zanin, and E. Menasalvas-Ruiz. Disnet: a framework for extracting phenotypic disease information from public Sources. PeerJ , 8(e8580), 2020. J. Pi  3. nero, J. M. Ram ırez-Anguita, J. Sa uch-Pitarch, F. Ronzano, E. Centeno, F. Sanz, and L. I. Furlong. The DisGeNET knowledge platform for disease genomics: 2019 update.Nucleic Acids Research 48(D1):D845–D855, 11 2019. 4. A. Saha, V. Pahuja, M. Khapra, K. Sankaranarayanan, and S. Chandar. Complex sequential question answering: Towards learning to converse over linked question answer pairs with a knowledge graph. Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 5. C. Sun et al. A natural language interface for querying graph databases PhD thesis, Massachusetts Institute of Technology, 2018. 6. L. Tian, X. Zhou, Y.-P. Wu, W.-T. Zhou, J.-H. Zhang, and T.-S. Zhang. Knowledge graph and knowledge reasoning: A systematic review. Journal of Electronic Science and Technology 20(2):100159, 2022. 21