Umme Afroz

Research

My research focuses on detecting adverse drug reactions (ADRs) in post-market pharmacovigilance using natural language processing and large language models.

Specifically, I investigate the following:

- Developing hybrid NLP pipelines that integrate traditional machine learning with LLM-based classifiers to extract adverse reactions from informal sources, comparing them against formal FDA reporting databases.

- Designing and validating statistical approaches with log-scaling and on-label use adjustments to improve signal detection.

- LLM-based screening pipelines.

- Age and sex-stratified analysis of adverse drug reactions.

Publications

1. Revisiting Disproportionality Metrics: Log-Scaled and On-Label Use–Adjusted Approaches for Post-Market ADR Detection [Accepted, American Medical Informatics Association 2026 Annual Symposium]

2. Comparing Adverse Drug Reaction Reporting Patterns for ADHD Medications Across Formal and
Informal Sources: A Study of FAERS and Reddit [Accepted, 2026 ASIS&T Annual Meeting]

3. Deep Learning Approaches for Protein Secondary Structure Prediction. IEEE Xplore. Available: https://ieeexplore.ieee.org/document/11021832

4. Protein Secondary Structure Prediction Using Attention-Based Fusion of Language Models (pp. 357–372). Springer. Available: https://link.springer.com/chapter/10.1007/978-3-032-15346-3_25