Background and Objective: Drug-drug interactions (DDIs) present major challenges in clinical care, impacting patient safety and therapeutic outcomes. Traditional detection methods, often based on curated databases, may miss newly emerging interactions or require frequent manual updates. Recent advances in natural language processing (NLP) and artificial intelligence (AI) offer novel tools for real-time, automated DDI detection directly from biomedical literature. This study introduces an AI-powered approach to identify DDIs, assign relevance scores to publications, and benchmark performance against conventional databases.

Methods: We developed an NLP pipeline incorporating sentence classification to detect safety and interaction-related content, and named entity recognition (NER) to extract drug-drug interaction types, their clinical impact, and patient features. The sentence classification module is based on a BERT architecture fine-tuned from microsoft/BiomedNLP-BiomedBERT-baseuncased-abstract-fulltext, and the NER module uses the open-source model d4data/biomedical-ner-all. A DDI relevance score was designed to quantify each article’s significance based on weighted parameters derived from model outputs, with weights informed by domain experts (title relevance: 50 points, DDI impact: 40 points, publication type: 10 points). The system was validated on a manually curated dataset. We applied ANOVA and sensitivity analyses, and evaluated performance using recall, precision, and F1-score, comparing our method to established DDI resources.

Key Content and Findings: The pipeline was validated on 7,751 articles across 15 query types. ROC analysis identified a relevance score threshold of 75%. Drug-family queries achieved the highest F1-score (0.872), followed by single-drug (0.731) and drug-drug queries (0.684). ANOVA and Tukey post-hoc analysis confirmed statistically significant differences in relevance scores between annotation groups (p < 0.05), with full stratification across all three groups for drug-drug family queries. Compared with manually compiled expert reference lists, our system retrieved substantially more relevant articles without requiring manual updates. 

Conclusion: AI-driven NLP approaches can enhance clinical decision support systems by providing comprehensive, evidence-based DDI detection and prioritization, offering scalable support for safer pharmacological decision-making.