MEDPrompt

Prompts, tools, chains and agents for healthcare using LLMs & FHIR. ✍️

About

MEDPrompt is a collection of prompts, tools, chains and agents for medical applications using LangChain and space using Agency. MEDPrompt also includes a collection of templates for using FHIR in LLM prompts (see below). The aim of MEDPrompt is to provide a conceptual framework and a set of tools for building healthcare applications using LLMs. Please read my Blog post. User contributions are highly appreciated!

Terminology

• Prompts are inputs or queries to LLMs that users can provide to elicit specific responses from a Large Language Model (LLM). Example: You are an AI assistant. Summarize this clinical document in 250 words
• Tools are functions used by agents for getting things done. Example: To find patient ID from name.
• Chains are tools that use LLM calls to get things done. Example: Answer a clinical question based on patient health record using RAG
• Agents uses an LLM to orchestrate Chains and Tools to acheive the overarching goal. Example: Answer a doctors question related to a patient. Find patient, get health record, generate embedding and generate answer
• Space is an Agency based abstraction for an environment for agents to communicate according to the actor model.

An example (Names correspond to the files in the repository)

Disclaimer:

This repository is not associated with the Medprompt method of prompting. In this generic repository, I will be trying to implement the method using langchain abstractions. Get in touch to share your thoughts via GitHub discussions. Please submit a PR with a link to the official implementation if any.

FHIR2Text -> Convert FHIR resources to plain text

This repository includes templates for converting FHIR resources into a text representation that can be injected into an LLM prompt. Only relevant information is extracted from the resource with simple transformations using helper functions. 🚒See this example usage.

FHIR2Calculator -> Calculate clinical scores from a FHIR Bundle (Work in progress)

Clinical calculators are tools that help healthcare professionals make medical decisions by providing them with quick and easy access to various medical formulas, scores, and algorithms. Calculations performed by LLMs are not reliable. FHIR2Calculator performs calculations on data fields extracted from a FHIR bundle and outputs the results as plain text that can be injected into LLM prompts.

Healthcare tools and chains for LangChain agents

Tools are functions that Langchain agents can use to carry out tasks based on the tool description. Example usage:

from medprompt.tools import FhirPatientSearchTool
tools = [FhirPatientSearchTool()]

Documentation is in progress. Any help will be highly appreciated.

Documentation & List of Templates

Usage

See Examples folder

1. Observation
2. FHIR Bundle More documentation and examples to follow..

Install

• medprompt for the core package
• embedding for using RAG with HuggingFace transformers

pip install medprompt
pip install medprompt[embedding]

Install Develop branch

pip install git+https://github.com/dermatologist/medprompt.git

OR
after cloning the repository
pip install -e .[embedding]

Import

from medprompt import MedPrompter
prompt = MedPrompter()
prompt.set_template(
    template_name="fhir_search_oai_chat_v1.json")

print(prompt.get_template_variables())

messages = prompt.generate_prompt(
    {"question": "Find Conditions for patient with first name John?"})

print(messages)

• Example hosting using LangServe

Give us a star ⭐️

If you find this project useful, give us a star. It helps others discover the project.

Contributing

• PR welcome
• Add templates in this folder as jinja2 or JSON.
• Follow the naming conventions in the folder.
• Add tools, chains and agents in the appropriately named folders.
• Add documentations here as a markdown file with the same name.
• Add a link in the index.md file.
• Please see CONTRIBUTING.md

Contributers

• Bell Eapen |
• My Blog post