Automated systems to identify relevant documents in product risk management

Area of Science:

• Pharmacovigilance
• Health Informatics
• Text Mining

Background:

• Product risk management relies on assessing health product risks and benefits.
• Primary literature is a key safety information source, especially for novel products.
• Vast literature requires efficient methods for identifying relevant safety data.

Purpose of the Study:

• To explore text mining for automating the identification of useful articles in product risk management.
• To improve the efficiency of literature searches in pharmacovigilance.

Main Methods:

• Computed term-frequency inverse document-frequency (TF-IDF) values for predictors from article titles and abstracts.
• Developed a general automated system using general predictors and tested its generalizability.
• Created specific automated systems using general and specific predictors with varying training set sizes.

Main Results:

• The general system achieved an AUC of 0.731, ranking 34.6% of useful articles in the top 10%.
• Specific automated systems required only 20 articles for development.
• Specific systems demonstrated superior prediction performance (AUC 0.748) compared to the general system.

Conclusions:

• Specific automated systems can be developed quickly and overcome subjectivity in defining useful articles.
• Implementing specific automated systems enhances the efficiency of health product risk management.