A collaborative large language model for drug analysis

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Pharmacology

Background:

• Large language models (LLMs) demonstrate human-level fluency but pose risks in healthcare due to factual inaccuracies (hallucinations).
• Ensuring traceability of information sources is crucial for clinical adoption of AI tools.

Purpose of the Study:

• To develop a knowledge-grounded collaborative LLM, DrugGPT, for accurate and evidence-based clinical decision-making.
• To address the limitations of generic LLMs in healthcare by enhancing factual accuracy and source traceability.

Main Methods:

• DrugGPT integrates diverse clinical-standard knowledge bases.
• A collaborative mechanism adaptively analyzes inquiries, identifies relevant knowledge, and aligns them for drug-related queries.
• Evaluated on drug/dosage recommendations, adverse reaction/drug-drug interaction identification, and pharmacology questions.

Main Results:

• DrugGPT demonstrated superior performance compared to existing LLMs across all evaluated metrics.
• Achieved state-of-the-art results with a reduced parameter count compared to generic LLMs.
• Ensured accurate, evidence-based, and faithful recommendations.

Conclusions:

• DrugGPT offers a reliable solution for clinical decision support by mitigating LLM hallucinations.
• The knowledge-grounded collaborative approach enhances the safety and trustworthiness of AI in pharmaceutical applications.