From UMLS to ClinicalBERT: Deep hybrid-health recommender system (Dh_HRS)

Background The digitalisation of healthcare and the development of clinical decision support systems represent a major step towards improving care quality and reducing medical errors. Objective To develop a hybrid medical recommendation system that delivers robust, interpretable predictions for diagnoses, prescriptions, and clinical notes to enhance clinical decision support. Methods We propose a hybrid framework that uses Unified Medical Language System (UMLS) ontology and Concept Unique Identifiers (CUIs) to standardise patient profiles. ClinicalBERT was employed to generate contextual embeddings capturing linguistic and semantic nuances. These embeddings were integrated into a knowledge graph and processed using a Graph Convolutional Network (GCN) to enable contextual aggregation of semantic neighbourhoods. The system was evaluated experimentally against recent benchmark models using the MIMIC-III, PubMed, and ICD-10-CM datasets. Results Experiments conducted on the benchmark datasets demonstrated that the proposed hybrid approach achieved superior performance compared with recent existing benchmark models. Conclusion Combining UMLS ontology, ClinicalBERT contextual embeddings, and GCN-driven knowledge graphs enables robust and interpretable clinical predictions, representing a significant advancement in clinical decision support systems.