kiran78487/Rag_Symptoms_checker

🏥 Medical RAG System - Symptom-to-Disease Diagnosis

A production-grade medical RAG (Retrieval-Augmented Generation) system that provides intelligent differential diagnosis from natural language symptom descriptions.

⚡ Quick Start

API Endpoint

POST /api/v1/diagnose

Example Request

{
  "text": "I have sharp chest pain radiating to my left arm with sweating"
}

Example Response

{
  "emergency_flag": true,
  "emergency_message": "🚨 EMERGENCY: Possible heart attack. Call 911 immediately!",
  "extracted_symptoms": [
    {
      "original": "sharp chest pain",
      "canonical": "chest pain",
      "confidence": 0.95
    }
  ],
  "top_differentials": [
    {
      "disease_name": "Myocardial infarction",
      "confidence": 0.82,
      "matched_symptoms": ["chest pain", "arm pain", "sweating"],
      "reasoning": "Classic presentation..."
    }
  ]
}

🏗️ Architecture

This system implements 8 intelligent layers:

1. 1.Frontend Layer - REST API for mobile/web apps
2. 2.LLM Extraction - Structured symptom extraction from natural language
3. 3.Symptom Mapping - Dictionary + FAISS semantic matching to canonical terms
4. 4.Disease Retrieval - FAISS vector search across 1,394 diseases
5. 5.Scoring Engine - Hybrid 5-component formula:
6. 6.40% Vector similarity
7. 7.25% Symptom overlap
8. 8.15% Coverage ratio
9. 9.15% Modifier matching (OPQRST)
10. 10.5% Risk adjustment
11. 11.Safety Layer - Emergency detection for life-threatening conditions
12. 12.LLM Explanation - Controlled reasoning generation
13. 13.Response Formatting - Structured JSON output

📊 Data & Models

Knowledge Base

• —1,394 diseases from authoritative sources (Mayo Clinic, NHS)
• —3,371 canonical symptoms with 15,000+ variant mappings
• —Pre-built FAISS indices for fast semantic search

AI Models

• —LLM: Qwen2.5-3B-reasoning-medical-symptoms
• —Embeddings: BAAI/bge-small-en-v1.5 (384 dimensions)
• —Vector Search: FAISS IndexFlatIP (exact search)

🚨 Safety Features

Emergency Detection

The system automatically detects and flags:

• —❤️ Heart attack symptoms (chest pain + radiation + sweating)
• —🧠 Stroke indicators (facial drooping, speech difficulty, numbness)
• —🫁 Severe respiratory distress
• —🩸 Severe bleeding/trauma
• —😵 Loss of consciousness
• —🦠 Sepsis and meningitis patterns

When emergencies are detected, the system immediately returns an emergency alert and bypasses normal disease ranking.

📡 API Documentation

Setup (HuggingFace Spaces)

1. 1.Add your HuggingFace API token as a Space secret:
2. 2.Go to your Space settings
3. 3.Navigate to "Variables and secrets"
4. 4.Add a new secret: HF_TOKEN = your_huggingface_token
5. 5.This is required for LLM inference API calls

1. 1.The Space will automatically build FAISS indices on first deployment

Endpoints

1. Diagnose Symptoms

POST /api/v1/diagnose

Main diagnostic endpoint for symptom analysis.

2. Search Symptoms

GET /api/v1/symptoms/search?query=headache&top_k=10

Semantic search for canonical symptoms (useful for autocomplete).

3. Health Check

GET /api/v1/health

Check if models are loaded and system is ready.

4. Interactive Docs

GET /docs

Swagger UI for testing the API.

🔧 Technical Details

Scoring Formula

FINAL_SCORE = 
  (0.40 × vector_similarity) +
  (0.25 × symptom_overlap) +
  (0.15 × coverage_ratio) +
  (0.15 × modifier_match) +
  (0.05 × risk_adjustment)

Symptom Matching

1. 1.Dictionary lookup - Instant exact matching (98% confidence)
2. 2.FAISS semantic search - Fuzzy matching with threshold ≥ 0.80
3. 3.Gap validation - Ensures top match is significantly better

Disease Retrieval

1. 1.Combine canonical symptoms into query
2. 2.Embed using BGE model (384D)
3. 3.Search FAISS index for Top-20 candidates
4. 4.Re-rank using hybrid scoring
5. 5.Return Top-5 with explanations

⚠️ Medical Disclaimer

CRITICAL: This system is designed for informational purposes only and is NOT a substitute for professional medical advice, diagnosis, or treatment.

• —❌ DO NOT use this for self-diagnosis
• —❌ DO NOT delay seeking medical care based on results
• —❌ DO NOT ignore emergency symptoms
• —✅ ALWAYS consult qualified healthcare professionals
• —✅ CALL 911 for emergencies indicated by the system

📱 Mobile App Integration

React Native Example

async function diagnose(symptoms) {
  const response = await fetch('https://barathvasan-rag-system-med-connect.hf.space/api/v1/diagnose', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({ text: symptoms })
  });
  return await response.json();
}

const result = await diagnose("I have severe headache with stiff neck");
if (result.emergency_flag) {
  alert(result.emergency_message);
}

Flutter Example

Future<Map> diagnose(String symptoms) async {
  final response = await http.post(
    Uri.parse('https://barathvasan-rag-system-med-connect.hf.space/api/v1/diagnose'),
    headers: {'Content-Type': 'application/json'},
    body: jsonEncode({'text': symptoms}),
  );
  return json.decode(response.body);
}

🧪 Testing

Test the API with sample symptoms:

curl -X POST "https://barathvasan-rag-system-med-connect.hf.space/api/v1/diagnose" \
  -H "Content-Type: application/json" \
  -d '{"text": "Sharp chest pain radiating to left arm with sweating and nausea"}'

📝 License

MIT License - See LICENSE file for details

👨‍💻 Developer

Built by Barathvasan

🙏 Acknowledgments

• —Medical data sources: Mayo Clinic, NHS
• —LLM: Qwen2.5-3B-reasoning-medical-symptoms by dumbequation
• —Embeddings: BGE by Beijing Academy of Artificial Intelligence
• —Vector search: FAISS by Meta AI

Remember: This is a research/educational tool. For medical concerns, always consult healthcare professionals.