AI Billing Rule Extraction Pipeline 🏥

Extract medical billing rules in formal, machine-usable form from natural-language billing manual PDFs.

Proof of Concept, code is messy 🔥

Demo

billing_rules.mp4

Run

You'll need an OpenAI API key: Copy .env.sample to a new file called .env and fill it in.

Then, you can start the application (database and rules extractor UI) with Docker Compose:

docker compose up

The rules extractor will be available at http://localhost:8501. You can either upload PDFs in the UI, or use the preconfigured NCCI ones. Then click "Run Extraction Analysis" to trigger the pipeline.

Pipeline Steps

1. Extraction: Parse the PDF into raw text (no AI/OCR needed for NCCI, but could be nice addition for more complex documents).
2. Use LLM with constrained generation to generate a flattened JSON AST. (Flattened representation avoids generation errors common with complex/recursive schemas).
3. Transform the generated flat-AST into a more standard (recursive) format and display it to the user.

The rules are displayed next to the PDF and extracted text in the rules extractor UI for easy human review.

Improvements

Currently, the system sometimes extracts text descriptions (e.g., "endoscopy") instead of standardized billing codes. To be production-ready, it needs integration with external classification databases e.g. to map procedures to CPT codes and diagnoses to ICD codes.

That task might be well-suited for an agentic setup with a web-search MCP, giving an extraction agent the discretion to look up codes when needed and edit the rules accordingly.

Rules Engine

A rules engine exists as well in this repository, but it's WIP.