https://github.com/cocoindex-io/patient-intake-extraction
Patient Intake Form Extraction using llm
https://github.com/cocoindex-io/patient-intake-extraction
data-indexing etl healthcare llm machine-learning ocr rag
Last synced: 9 days ago
JSON representation
Patient Intake Form Extraction using llm
- Host: GitHub
- URL: https://github.com/cocoindex-io/patient-intake-extraction
- Owner: cocoindex-io
- License: apache-2.0
- Created: 2025-03-21T03:47:21.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2026-06-11T05:59:35.000Z (about 1 month ago)
- Last Synced: 2026-06-21T06:24:51.660Z (about 1 month ago)
- Topics: data-indexing, etl, healthcare, llm, machine-learning, ocr, rag
- Language: Python
- Homepage: https://cocoindex.io
- Size: 649 KB
- Stars: 16
- Watchers: 2
- Forks: 3
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# 🥥 Patient Intake Extraction with CocoIndex
Extract structured data from patient intake forms (PDF) using [CocoIndex](https://cocoindex.io) v1 and [DSPy](https://github.com/stanfordnlp/dspy) with a Gemini vision model. Each form is rendered to images, passed to the LLM, and written out as a validated JSON record — incrementally, so re-runs only reprocess forms that actually changed.
❤️ Please give [CocoIndex on GitHub](https://github.com/cocoindex-io/cocoindex) a star ⭐ to support us. A warm coconut hug 🥥🤗. [](https://github.com/cocoindex-io/cocoindex)
## What it does
```
data/patient_forms/*.pdf ──▶ pages → images ──▶ DSPy + Gemini ──▶ output_patients/*.json
```
- **Pydantic models** (`models.py`) define the target schema — type-safe, validated `Patient` records.
- **DSPy module** (`main.py`) extracts directly from page images via `ChainOfThought` with vision — no intermediate text/markdown step.
- **CocoIndex v1 app** (`main.py`) processes each form as an independent component and writes JSON, recomputing only what changed.
## Prerequisites
- **Python 3.11+**
- A **Gemini API key** — no Postgres, no Docker.
## Run
**1. Install dependencies** (sample PDFs are already in `data/patient_forms/`):
```bash
pip install -e .
```
**2. Set your API key** — copy the template and add your Gemini key:
```bash
cp .env.example .env
# then edit .env and set GEMINI_API_KEY=your_api_key_here
```
The template also sets `COCOINDEX_DB=./cocoindex.db` (the local engine state path), which is required.
**3. Extract:**
```bash
cocoindex update main.py
```
This reads each PDF in `data/patient_forms/`, extracts the patient record, and writes JSON to `output_patients/`:
```bash
ls output_patients/
# Patient_Intake_Form_David_Artificial.json
# Patient_Intake_Form_Emily_Artificial.json
# Patient_Intake_Form_Joe_Artificial.json
# Patient_Intake_From_Jane_Artificial.json
```
## Incremental processing
Re-running never redoes work that's already done — `process_patient_form` is memoized by file content:
```bash
cocoindex update main.py # ⚡ unchanged forms are skipped
```
Add, replace, or remove a PDF and re-run — only the affected form is reprocessed, and a removed PDF's JSON is cleaned up automatically.
## How it works
The pipeline lives in [`main.py`](./main.py):
- `extract_patient` — a `@coco.fn` that renders each PDF page to an image (pymupdf) and runs a DSPy `ChainOfThought` vision module to return a validated `Patient` ([`models.py`](./models.py)).
- `process_patient_form` — a memoized `@coco.fn` that reads a file, extracts, and declares the JSON output.
- `app_main` — walks `data/patient_forms/` and mounts one component per PDF.
Each form is processed independently and memoized by content, so re-runs only touch what changed.
---
The sample forms under `data/` are purely artificial and for testing only.