mou11/medical-report-generator

Medical Report Generator

Live Demo

👉 Try it on Hugging Face Spaces

Overview

An end-to-end clinical report generation pipeline that generates structured medical reports from patient data, detects hallucinations using NLI, outputs FHIR R4 compliant JSON, and exports professional PDF reports.

Results

Report Type	BERTScore F1	Safety Score
Radiology	0.8628	0.625
Discharge Summary	0.9045	1.0
Lab Report	0.8129	0.375

Hallucination Detection

Report Type	Total Claims	Hallucination Rate	Safety Score
Radiology	4	0.375	0.625
Discharge	1	0.0	1.0
Lab	4	0.625	0.375

Features

• Generates 3 types of clinical reports: Radiology, Discharge Summary, Lab Report
• Hallucination detection using NLI (cross-encoder/nli-deberta-v3-base)
• FHIR R4 compliant JSON output (HL7 healthcare standard)
• BERTScore and ROUGE evaluation metrics
• Professional PDF export with quality assessment table
• Gradio web interface

Architecture

Patient Data → Report Generation (LLaMA 3.3 70B via Groq) → Hallucination Detection (DeBERTa NLI) → Evaluation (BERTScore + ROUGE) → FHIR R4 JSON → PDF Export

Tech Stack

Component	Tool
LLM	Groq — LLaMA 3.3 70B
Hallucination Detection	cross-encoder/nli-deberta-v3-base
Evaluation	BERTScore + ROUGE
FHIR Output	HL7 R4 JSON
PDF Export	ReportLab
UI	Gradio
Platform	Google Colab (T4 GPU)

How to Run

1. Get a free Groq API key at console.groq.com
2. Open app.py in Google Colab
3. Add your key to Colab Secrets as GROQ_KEY_1
4. Run all cells in order

Medical Disclaimer

This system is for educational and research purposes only. It does not provide medical advice. Always consult a qualified healthcare professional for medical decisions.

Project Status

✅ Report generation pipeline complete ✅ Hallucination detection implemented ✅ FHIR R4 compliant output ✅ PDF export working ✅ Gradio demo deployed on Hugging Face Spaces