TL;DR
This paper introduces MedAssist, an LLM-powered system designed to help clinicians and patients understand complex Electronic Health Records (EHRs). By integrating external medical knowledge retrieval with internal EHR data analysis, MedAssist aims to streamline clinical decision-making and improve patient care.
What Problem Does This Paper Address?
Comprehending diagnoses and treatment plans, especially for rare diseases or complex patient histories, is a significant challenge for both medical professionals and patients. Electronic Health Records (EHRs) contain a wealth of information, but extracting actionable insights and contextualizing them with up-to-date medical knowledge is time-consuming and requires specialized data engineering skills. Existing LLM applications in healthcare often focus on specific tasks like disease prediction or question answering, but there's a need for a unified framework that systematically integrates LLMs into broader EHR workflows to support comprehensive clinical understanding and decision-making.
Methods
MedAssist is built around a modular framework that combines two key capabilities:
- Knowledge Retrieval from Medical Corpora: This module leverages a diverse external corpus (PubMed, DrugBank, MeSH, Wikipedia, MS MARCO) to retrieve relevant medical knowledge. It uses a BMRetriever model for efficient retrieval and a cross-encoder for reranking passages to improve quality. This step injects up-to-date medical information without requiring expensive model retraining.
- Data Retrieval from Internal Records: This module utilizes an "EHRAgent" for multi-turn interactive coding with external tools. It decomposes complex queries into manageable actions, enabling interaction with EHR databases. MedAssist integrates several tools for this purpose:
- Database Loader: Loads specific tables.
- Data Filter: Filters tables based on conditions.
- Get Value: Retrieves column values or performs basic operations.
- Calculator: Performs calculations using the WolframAlpha API.
- Calendar: Computes dates.
- SQL Interpreter: Executes SQL queries generated by LLMs.
These components work together, with LLMs acting as autonomous agents that iteratively refine code based on execution feedback until an optimal solution is reached. The system then uses a Knowledge Summarization with LLMs module to generate personalized, context-aware patient summaries by integrating information from both internal EHR data and external knowledge sources.
Datasets
The paper mentions using publicly available EHR data, specifically citing MIMIC as an example. For external knowledge, they curate a diverse corpus comprising:
- PubMed
- DrugBank
- MeSH
- Wikipedia
- MS MARCO
The paper does not introduce a new benchmark dataset but rather focuses on integrating existing data sources.
Key Insights and Ideas
- Bridging the EHR-Knowledge Gap: The core innovation is the seamless integration of external, up-to-date medical knowledge with internal patient EHR data. This addresses a critical gap in current clinical practice.
- LLMs as Clinical Assistants: MedAssist demonstrates how LLMs can be effectively leveraged not just for information retrieval but also for complex reasoning, data manipulation, and generating human-readable summaries, acting as a true assistant to clinicians.
- Tool Use for EHR Interaction: The framework's ability to integrate and utilize various external tools (like SQL interpreters and calculators) allows LLMs to perform sophisticated operations on structured EHR data, overcoming limitations of pure text-based processing.
- Iterative Refinement for Accuracy: The multi-turn conversation between the LLM and the code executor, with feedback loops for error correction, is crucial for achieving accurate and reliable results from complex EHR queries.
- User-Friendly Interfaces: The emphasis on user-friendly interfaces tailored for both clinicians and patients is key to the practical adoption and utility of such a system.
Key Results
The paper presents a case study of a 14-year-old patient with adolescent idiopathic scoliosis and mild asthma. MedAssist successfully:
- Retrieved literature supporting bracing and physiotherapy for scoliosis, noting no adverse respiratory effects.
- Queried the EHR database to find 22 similar patient profiles, identifying treatments and outcomes.
- Generated a concise summary emphasizing that the proposed treatment plan aligns with established practices, recommending specific protocol requirements (e.g., 16 hours of bracing per day) and structured follow-up evaluations.
The system demonstrated its ability to provide informative and actionable summaries, highlighting diagnostic insights and clinical considerations.
Limitations and Future Work
The authors acknowledge the following:
- Current Data Sources: MedAssist currently uses online biomedical literature and publicly available EHR data (like MIMIC). Further adoption of data from other resources will require proper IRB approvals.
- Future Work: The authors plan to expand MedAssist's capabilities for multi-modal data, such as imaging and genomic data. They also aim to ensure its robustness across diverse patient populations. The core methods of the system will be open-sourced to allow customization within local clinical environments.
Summary updated