AI-powered prescription discrepancy detection system using OCR, LLMs, RxNorm, and RAG to identify clinical errors in prescriptions.
概览
RxDetect catches errors in handwritten and printed prescriptions before they reach the patient — combining computer vision OCR, deterministic clinical rules, real-time drug validation APIs, RAG-augmented clinical guidelines, and LLM reasoning into a single audit-ready workflow. RxDetect automates the first line of clinical safety review — a pharmacist's check that a prescription is complete, consistent, and correctly dosed. A clinician uploads a prescription image or PDF; the system runs it through a six-step AI pipeline and returns a structured discrepancy report with: - A classification label and confidence score - Per-rule findings (rule ID, severity, description) - A 27-parameter clinical audit checklist (YES / NO / N/A) - Clinical evidence retrieved from the institution's guideline PDFs - A pharmacist feedback form to correct the AI and close the loop Each uploaded prescription is processed asynchronously by a Celery worker through a six-step pipeline: The frontend was...
README
以下为原文 README 节选，完整内容请查看原文。
RxDetect
AI-powered prescription discrepancy detection. RxDetect catches errors in handwritten and printed prescriptions before they reach the patient — combining computer vision OCR, deterministic clinical rules, real-time drug validation APIs, RAG-augmented clinical guidelines, and LLM reasoning into a single audit-ready workflow.
Table of Contents
- Overview
- How It Works
- Discrepancy Classes
- Tech Stack
- Project Structure
- Getting Started
- Ingesting Clinical PDFs into the Knowledge Base
- API Reference
- Frontend Pages
- Contributing
- License
Overview
RxDetect automates the first line of clinical safety review — a pharmacist’s check that a prescription is complete, consistent, and correctly dosed. A clinician uploads a prescription image or PDF; the system runs it through a six-step AI pipeline and returns a structured discrepancy report with:
- A classification label and confidence score
- Per-rule findings (rule ID, severity, description)
- A 27-parameter clinical audit checklist (YES / NO / N/A)
- Clinical evidence retrieved from the institution’s guideline PDFs
- A pharmacist feedback form to correct the AI and close the loop
Key capabilities
| Capability | Details |
|---|---|
| OCR extraction | GPT-4o Vision extracts all prescription fields from any image or PDF format |
| 27-parameter checklist | A second GPT-4o call scores 27 clinical parameters (YES / NO / N/A) per prescription |
| Deterministic rule engine | 20+ clinical rules covering dose range, duplicate drugs, allergy flags, missing fields, and illegibility |
| Drug validation | RxNorm + OpenFDA real-time cross-check of drug names, dosage, interactions, and contraindications |
| RAG evidence retrieval | FAISS vector store seeded with your institution’s clinical guideline PDFs |
| LLM reasoning | GPT-4o synthesises all signals into a final verdict, confidence score, and clinical narrative |
| PDF audit reports | ReportLab generates downloadable, audit-ready PDFs — generated in-memory, never written to disk |
| Pharmacist feedback loop | Pharmacists can correct AI verdicts; feedback is stored for future model improvement |
| Role-based access | JWT access + refresh tokens; pharmacist / admin / viewer roles |
| Prescription comparison | Side-by-side 27-parameter checklist comparison across any number of analysed prescriptions |
How It Works
Each uploaded prescription is processed asynchronously by a Celery worker through a six-step pipeline:
Upload → [Celery queue]
│
├── Step 1: OCR (GPT-4o Vision)推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add harshithgr5/rxdetect