MedPlan:A Two-Stage RAG-Based System for Personalized Medical Plan Generation

2025-03-23

AI Summary

Existing EHR-based language models (EHR-LMs) predominantly focus on clinical assessment rather than treatment planning, exhibiting three key limitations: (1) insufficient clinical sequential reasoning, (2) neglect of longitudinal patient history, and (3) conflation of subjective and objective clinical information. To address these, we propose a two-stage Retrieval-Augmented Generation (RAG) framework aligned with the SOAP (Subjective, Objective, Assessment, Plan) clinical workflow: Stage 1 generates a clinical assessment from chief complaint and objective data; Stage 2 produces a personalized, structured treatment plan by integrating retrieved historical EHR records. Our contributions are threefold: (1) the first workflow-aware, staged clinical reasoning paradigm; (2) explicit modeling of subjective–objective information separation; and (3) deep integration of temporal EHR context. Experiments demonstrate significant improvements over baselines in assessment accuracy, as well as clinical plausibility and personalization of generated treatment plans.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: Generation
• Knowledge Representation and Reasoning: Knowledge Representation Languages

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Despite recent success in applying large language models (LLMs) to electronic health records (EHR), most systems focus primarily on assessment rather than treatment planning. We identify three critical limitations in current approaches: they generate treatment plans in a single pass rather than following the sequential reasoning process used by clinicians; they rarely incorporate patient-specific historical context; and they fail to effectively distinguish between subjective and objective clinical information. Motivated by the SOAP methodology (Subjective, Objective, Assessment, Plan), we introduce MedPlan, a novel framework that structures LLM reasoning to align with real-life clinician workflows. Our approach employs a two-stage architecture that first generates a clinical assessment based on patient symptoms and objective data, then formulates a structured treatment plan informed by this assessment and enriched with patient-specific information through retrieval-augmented generation. Comprehensive evaluation demonstrates that our method significantly outperforms baseline approaches in both assessment accuracy and treatment plan quality.

Problem

• Generates treatment plans using sequential reasoning like clinicians
• Incorporates patient-specific historical context for personalized plans
• Distinguishes subjective and objective clinical information effectively

Innovation

• Two-stage RAG architecture for medical plans
• Aligns LLM reasoning with clinician workflows
• Incorporates patient-specific historical context