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

TL;DR

MedPlan tackles the gap where prior LLM-based EHR systems focus mainly on assessment, by introducing a SOAP-driven two-stage framework that first generates an assessment from $S$ and $O$, then formulates a patient-specific plan $P$ augmented with longitudinal context via retrieval streams. The approach employs two specialized LLMs for $A$ and $P$, retrieval-augmented generation with self-history and cross-patient references, and instruction-tuning to align outputs with clinical reasoning. Evaluation on a large, de-identified FEMH SOAP-note dataset shows consistent improvements in assessment and plan quality over single-pass baselines across multiple backbones, plus a clinical prototype indicating real-world viability. These results suggest that structuring LLM reasoning along the clinician’s SOAP workflow and leveraging long-context, cross-patient information can meaningfully enhance medical planning in EHR environments.

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 \ours{}, 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.

Figures (6)

• Figure 1: Compare the existing approach (left) with our proposed MedPlan (right). We adopt the SOAP protocol and simulate the doctor diagnosis process with LLM for medical plan generation.
• Figure 2: Overall architecture of the proposed MedPlan framework.
• Figure 3: Plan Generation Results: Human Doctor, Baseline LLM, and MedPlan
• Figure 4: Overview of the Clinical Application of the MedPlan System
• Figure 5: MedPlan System Architecture.