MediPhen: Prompt-Based LLM Reasoning with Synthesized Multimodal Clinical Knowledge for Zero-Shot Multi-morbidity...

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Clinical Decision Support

Background:

• Clinical decision support systems leverage electronic health records (EHRs) for disease classification.
• Large language models (LLMs) excel at processing unstructured clinical notes but struggle with multimodal data integration.
• A unified framework is needed to combine structured lab results and unstructured notes for zero-shot clinical decision support.

Purpose of the Study:

• To introduce MediPhen, a novel reasoning framework for multimodal disease phenotyping using LLMs.
• To adapt LLMs for zero-shot multi-morbidity phenotyping by integrating structured and unstructured EHR data.
• To enhance LLM transfer learning and clinical reasoning through knowledgebase integration and chain-of-thought prompting.

Main Methods:

• MediPhen adapts LLMs for zero-shot phenotyping using extracted clinical entities, relations, and lab narratives.
• A clinical knowledgebase guides phenotype classification and improves LLM transfer learning.
• An explanation module uses chain-of-thought prompting for enhanced clinical reasoning.

Main Results:

• Experiments on MIMIC-III and MIMIC-IV benchmarks demonstrate MediPhen's effectiveness across multiple LLMs.
• MedGemma-27B achieved state-of-the-art results, improving F1 scores by 19.92% (MIMIC-III) and 16.68% (MIMIC-IV) over fine-tuned baselines.
• MediPhen effectively integrates structured and unstructured EHR data for advanced clinical AI.

Conclusions:

• MediPhen serves as a scalable, zero-shot screening tool for multi-morbidity phenotype classification.
• The framework advances the integration of diverse EHR data types for clinical decision support.
• MediPhen demonstrates the potential of LLMs in multimodal clinical data analysis and reasoning.