Accelerating dataset generation for machine learning using large language models: a pharmaceutical additive...

Area of Science:

• Pharmaceutical research
• Machine learning
• Data science

Background:

• High-quality datasets are crucial for machine learning (ML) in specialized fields like pharmaceutical research.
• Manual data extraction from scientific literature is time-consuming and labor-intensive.
• Existing methods struggle with heterogeneous data sources in drug development.

Purpose of the Study:

• To develop a novel deep prompt-engineering framework to automate dataset generation for pharmaceutical research.
• To transform GPT-4 into a tool for accelerated and accurate extraction of critical parameters from literature.
• To reduce the manual effort and time required for creating structured datasets for ML models.

Main Methods:

• A multi-set prompt strategy was employed using GPT-4 to analyze 70 full-text pharmaceutical inkjet printing articles.
• 22 domain-relevant variables were extracted and computed, categorized into printing, rheological, and drug dose parameters.
• Outputs were rigorously benchmarked against a human-curated dataset compiled by domain experts.

Main Results:

• The GPT-4 framework achieved an overall accuracy of 0.942 across 4,217 data points.
• Computed variables demonstrated high accuracy (0.983), even with complex calculations and unit conversions.
• Processing time per article was reduced from hours of human effort to under 3.5 minutes.

Conclusions:

• The novel prompt-engineering approach enables GPT-4 to generate reliable, high-quality, literature-derived datasets.
• This method significantly reduces manual effort while maintaining expert-level accuracy for ML model training.
• The strategy facilitates the scalability of machine learning in pharmaceutical and other data-intensive domains.