/LLM-Medical-Analyser

Primary LanguageJupyter Notebook

Certainly! Here's a basic README template for your Python app:

Medical Doc Analyser

This Python app is designed to Ingest and Analyse medical records and approve them for treatment.

How It Works

  • experiments is notebook for different approaches tried

  • app.py contains the main logic

  • simple_app.py contains the naive logic to get the prototype working

  • utils.py continas helper functions

  • llm.py contains code for interacting with models

  • task directory contains the medical data

  • Step 1: get the doc structured

  • Step 2: Determine the treatment that is recomended by the doctor

  • Step 3: Determine is prior treatments improved symptoms

  • Step 4: If Not: Ingest the Enriched Guidelines + Examples

  • Step 5: Check if Patient Qualifies Or Suggest the additional information needed : Judge

Design Decisions:

  • Decision 1: Use RAG - May work, able to extract relevant sections but need further work to add to pipeline
  • Decision 2: Use Multiple Models: Multiple LLMs (currently claude + openai) allow us to avoid potential hallucinations and get higher confidence of decisions
  • Decision 3: Use LLMS to structure the document: LLMs are quite good at parsing documents, since the content is similar we can use LLMs to create objects
  • Decision 4: Chain of Thought (COT) reasoning when possible, research has shown models do better with COT Decision 5: 1 shot example for final answer.

Future Plans

  • The following are some short comings / remaining improvements that could be addressed given more time

  • Use a model fine tuned on clinical data

  • Improve RAG to fetch relevant Data (See experiments.ipynb) for current state

  • Use APE, Automated Prompt Engineering based on a reward for the model when the right output is returned

  • Scale -> Dockerize + Plug to AWS Lambda - Use Opensearch / Elastic Search for RAG

Getting Started

Prerequisites

  • Conda
  • OpenAI API Key
  • (Optional) For Judging LLMs, using Claude, AWS Setup. Please ensure claude in bedrock is enabled in the account, the region is us-east-1.

Installation

  1. Clone the repository:

    git clone https://github.com/balajsaleem/MedicalAnalyser.git

  2. Navigate to the project directory:

    cd MedicalAnalyser

  3. Create a Conda environment:

    conda create --name your-environment-name python=3.10

  4. Activate the Conda environment:

    conda activate your-environment-name

  5. Install the required dependencies:

    pip install -r requirements.txt

Usage

Run the main application:

python app.py --medical_record_path=task/medical-record-3.pdf

Arguments

  • medical_record_path: The path to the medical record to ingest - string
  • judge_llms: Flag to enable LLM output pooling + judgement