Mapping Free-Text Symptoms to Medical Concepts Using Knowledge Graph Embeddings and Natural Language Inference

Sammanfattning

Artificial Intelligence (AI) is increasingly applied in healthcare to support clinical decision-making and improve patient outcomes. A major challenge is that patients often describe their symptoms using informal and ambiguous language that does not align with standardized medical vocabularies. As a result, mapping patient-written symptom descriptions to structured medical ontologies remains difficult, limiting the ability of AI systems to fully incorporate patient input. This thesis explores an approach that combines Knowledge Graph Embeddings (KGE) with Natural Language Inference (NLI) to position free-text symptoms in the same embedding space as standardized medical concepts. Using a RotatE-based KGE model trained on selected vocabularies from the Unified Medical Language System (UMLS) and entailment scores from a multilingual NLI model, we generated vector representations for free-text symptom descriptions. In evaluation using 231 real-world chief complaints from the MIMIC-IV-ED dataset and 66 manually annotated SNOMED CT concepts from UMLS, our method achieved a Top-10 accuracy of 74.7%, showing good alignment with expert annotations. The embeddings placed specific symptoms close to their relevant medical concepts, and the results demonstrated that ontology-aware embeddings, combined with semantic inference, outperformed NLI alone, leading to more accurate and meaningful mappings. This proof of concept highlights a promising approach for bridging patient language and standardized medical knowledge. It can support the development of clinical decision support tools and patient-facing applications, and with future extensions to larger datasets, multilingual settings, and uncertainty estimation, the approach shows strong potential for real-world applicability and further exploration.