Clinical large language model centered on electronic medical records

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Natural Language Processing in Healthcare

Background:

• Existing large-language models (LLMs) struggle to replicate the nuanced decision-making of medical professionals, especially in complex diagnostic scenarios.
• There is a need for AI tools that can bridge the gap between artificial intelligence and real-world clinical practice.
• Current AI platforms lack the ability to fully integrate diverse data sources like electronic medical records (EMR) and physician expertise.

Purpose of the Study:

• To introduce AI4Doctor, a novel LLM specifically designed for the clinical domain.
• To develop an innovative integration strategy for LLMs using distilled EMR data and physician insights.
• To create a more realistic medical practice environment for AI-driven consultation support.

Main Methods:

• Supervised fine-tuning of an LLM using a combination of distilled EMR data and empirical insights from practicing physicians.
• Implementation of a curriculum learning approach to manage the merging of diverse instructional sources during fine-tuning.
• Development of a novel reinforcement-learning approach with a reward system to align LLM outputs with clinical expertise (diagnostic priors, risk thresholds, heuristic saliencies).

Main Results:

• AI4Doctor demonstrates an enhanced ability to mimic the clinical acumen of healthcare practitioners.
• The hybrid model shows potential in improving diagnostic accuracy and decision-making in complex medical scenarios.
• Comparative evaluation using a subjective system with expert assessment indicates promising performance.

Conclusions:

• The AI4Doctor model, by synergizing EMR data and physician expertise, offers a robust tool for medical consultations.
• This hybrid approach effectively bridges the gap between AI capabilities and the complexities of clinical practice.
• The study highlights the potential of advanced LLMs, trained on integrated data, to augment healthcare professional decision-making.