Compositional Knowledge Graph for Clinical Decision Support in Pediatric Allergy Diagnostics

Allergies are immune system reactions to typically harmless substances, and diagnosing them in children is especially challenging. Moreover, the unstructured and non-standardized nature of clinical data hinders both clinical decision-making and the effectiveness of automated decision support systems. We present a compositional knowledge graph to analyze unstructured pediatric allergy data. We exploited a Large Language Model with LangChain to extract clinical entities, which were then organized into a Neo4j knowledge graph representing patients, allergens, and symptoms. Cypher queries were used to identify prevalent allergens, symptom–allergen associations, and predict unseen allergies based on symptom patterns. The results show that dust mites were the most common allergens, followed by pollen and food allergens. Respiratory problems were strongly associated with environmental allergens, while gastrointestinal symptoms were associated with food allergens. Network analysis revealed key allergen hubs and co-occurring allergen patterns, highlighting the graph’s potential for decision support. This study demonstrates how knowledge graphs can enhance pediatric allergy diagnosis by providing transparent insights for uncovering complex allergen–symptom relationships. It offers a promising approach for bridging unstructured data with clinical decision-making, paving the way for enhanced real-world applications in medical Artificial Intelligence.