MedReview: A Graph-Enhanced Framework for Comparing Domain-Specific and General Language Models in Consumer Drug Review Analysis

Venreddy, Ramakrishna Reddy (2025) MedReview: A Graph-Enhanced Framework for Comparing Domain-Specific and General Language Models in Consumer Drug Review Analysis. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (496kB) | Preview

Abstract

Consumer drug reviews represent a significant valuable source of pharmacovigilance information but present significant challenges due to the terminology gap between informal patient language and professional medical vocabulary. This paper introduces MedReview as one new framework making use of graph-based semantic bridging with language models to enhance the analysis of consumer drug reviews. The study addresses the research question: How can graph-based semantic bridging enhance the performance of both domain-specific and general language models in analyzing consumer drug reviews? The framework combines medical knowledge graphs constructed from UMLS and BioPortal with six language models: domain-specific models (PubMedBERT, BioBERT, ClinicalBERT) and general-purpose models (BERT, RoBERTa, DeBERTa). Through systematic experimentation on 50,000 drug reviews, the research demonstrates that graph enhancement provides substantial performance improvements across all models, with medical models showing the highest gains (22.39% average F1 improvement in sentiment classification). The best baseline model (DeBERTa) achieved an F1 score of 0.7042, while its graph-enhanced variant reached 0.8099, representing a 15% improvement. The multi-task assessment with sentiment classification, effectiveness regression, and multi-aspect classification proves fine-grained semantic tasks benefit the most with structured knowledge integration (average of 38.5% improvement). The work proves the graph-based semantic bridging across differences of vocabulary to be successful and enables the general and the medical model to capture the language of consumer health more effectively. This work contributes to healthcare informatics with the efficient process for bridging the gap between the healthcare systems and the patients and it can have real-world applications where pharmacovigilance and patient-centric care are involved.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Basilio, Jorge UNSPECIFIED
Subjects:	R Medicine > RS Pharmacy and materia medica P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing R Medicine > Healthcare Industry H Social Sciences > HM Sociology > Information Science > Communication > Medical Informatics
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	03 Jul 2026 11:33
Last Modified:	03 Jul 2026 11:33
URI:	https://norma.ncirl.ie/id/eprint/9467

Actions (login required)

View Item