Automatic Classification of Structured Product Labels for Pregnancy Risk Drug Categories, a Machine Learning Approach

Area of Science:

• Pharmacovigilance
• Computational drug safety

Background:

• Accurate drug risk information for pregnant patients is crucial for safe medication use.
• FDA-approved drug labels are a primary source for this information.

Purpose of the Study:

• To determine the prevalence of standard pregnancy drug risk categories in FDA-approved drug labels.
• To evaluate the feasibility of using machine learning to identify and suggest missing risk information.

Main Methods:

• 18,342 FDA drug labels were analyzed using regular expressions and manual review.
• A Sequential Minimal Optimization algorithm was trained on labeled data to classify pregnancy risk categories.
• Performance was evaluated using Micromedex categories, with a focus on document sections.

Main Results:

• 83% of analyzed drug labels contained explicit pregnancy risk categories.
• The 'precautions' section demonstrated the highest performance for risk category assignment (Accuracy: 0.79).
• Machine learning models achieved moderate performance in identifying existing risk information.

Conclusions:

• A significant portion of drug labels include pregnancy risk information, but 17% lack it.
• Machine learning, particularly models trained on the 'precautions' section, shows promise for identifying and potentially suggesting missing pregnancy risk data.
• Further development of AI tools could enhance the completeness of drug safety information for pregnant individuals.