A Hybrid Neural Collaborative Filtering and Word Embedding Approach for Personalised Diabetes Drug Recommendation

Personalised diabetes drug recommender systems have the potential to enhance treatment efficacy and advance precision medicine. However, their real-world implementation remains challenging due to issues such as data sparsity and the cold-start problem, which affect both new patients and newly introduced drugs. In this study, we propose a hybrid recommendation framework that integrates Neural Collaborative Filtering (NCF) with biomedical word-embedding techniques to capture complex patient–drug interactions while alleviating cold-start limitations. The NCF component learns nonlinear relationships from structured clinical variables and historical medication records, whereas the word-embedding component leverages BioBERT to extract semantic representations of drugs from textual descriptions, enabling effective retrieval of appropriate treatments even when interaction data is unavailable. Our approach was evaluated using the UCI Diabetes 130-US Hospitals dataset and a curated DrugBank corpus comprising 71 anti-diabetic drugs. The optimised NCF model achieved strong predictive performance, with an accuracy of 0.9004, precision of 0.7658, recall of 0.4209, F1-score of 0.5433, specificity of 0.9789, and AUC of 0.8717 on held-out data. Furthermore, the BioBERT-based semantic module generated clinically coherent drug similarity rankings, suitable for recommending alternatives in cold-start scenarios. By fusing patient- centric probability estimates with semantic similarity scores, the proposed hybrid strategy delivers ranked drug recommendations, supporting personalised and data-efficient diabetes treatment.