AI-Assisted Patient Registration System for Rural Indian Hospital

Challenge

A Hindu Mission Hospital in central Tamil Nadu serving underserved populations faced significant challenges with their patient registration process:

• Long Wait Times: Manual registration took 15-20 minutes per patient
• Duplicate Records: Elderly patients registering with family members' phone numbers created fragmented medical records
• Language Barriers: Staff struggled with varying literacy levels and language preferences
• Data Entry Errors: Manual entry error rate of 22%
• Limited Resources: Small staff couldn't efficiently handle patient volume during mobile health camps
• Infrastructure Challenges: Unreliable power supply and poor internet connectivity
• Paper Records: Difficult to maintain, search, and share patient data across village camps

Solution

We developed an AI-assisted patient registration system with human-in-the-loop verification, tailored for rural healthcare settings in India:

Face Recognition with Staff Verification

Multi-Setting Face Capture:

• Mobile app captures patient photos during registration and check-in
• AI suggests potential matches with existing patient records
• Staff reviews and confirms all AI-suggested matches
• Final approval always required from registration clerk
• Prevents duplicate registrations for patients who come with different family members
• Works in various lighting conditions (outdoor camps, clinic rooms)

Human-in-the-Loop Process:

// AI suggests matches, staff decides
const processPatientRegistration = async (faceImage, documentImage) => {
  // AI analyzes face
  const faceMatches = await faceRecognition.findSimilar(faceImage);

  // AI extracts document data
  const ocrResult = await aiOCR.process(documentImage);

  // Present to staff for verification
  return {
    suggestedMatches: faceMatches,  // Staff reviews these
    extractedData: {
      firstName: ocrResult.extract('FIRST_NAME'),
      lastName: ocrResult.extract('LAST_NAME'),
      dob: ocrResult.extract('DATE_OF_BIRTH'),
      address: ocrResult.extract('ADDRESS'),
      confidence: ocrResult.confidenceScores  // Highlight low-confidence fields
    }
  };
  // Staff verifies all data before final submission
};

Intelligent Document Processing with Validation

AI-Assisted OCR:

• Automatically extracts data from Aadhaar cards, ration cards, and insurance documents
• OCR highlights low-confidence fields for manual staff review
• Staff verifies and corrects all AI-extracted information
• Smart validation suggests corrections but requires staff approval
• Mobile app captures and processes documents instantly

Confidence-Based Review:

• Fields with <85% confidence flagged for manual verification
• Staff can easily edit any AI-suggested data
• All critical fields (name, DOB, address) require staff confirmation
• Visual highlighting of uncertain extractions

Voice-Enabled Interface with Bilingual Support

• Conversational prompts: Guides patients through registration in Tamil or English
• Speech-to-Text: Accurate transcription for basic information collection
• Accessibility: Designed for patients with limited literacy
• Staff reviews all voice-captured data before finalizing

Implementation

Phase 1: On-Site Assessment (3 weeks)

• Conducted workflow analysis at main hospital and 3 village camps
• Interviewed staff, community health workers, and patients
• Tested face recognition with diverse patient populations
• Assessed internet connectivity and power reliability
• Designed offline-first architecture for intermittent connectivity

Phase 2: Development (8 weeks)

• Built React Native mobile app for staff Android tablets
• Developed Python backend with TensorFlow for AI processing
• Implemented face recognition model trained on diverse Indian populations
• Created OCR confidence scoring system
• Integrated with existing paper-based workflow
• Designed offline queue with sync when connectivity available

Phase 3: Staff Training & Pilot (4 weeks)

• Trained 12 registration staff and community health workers
• Emphasized that AI suggestions require verification
• Conducted pilot at 2 village health camps with 156 patients
• Gathered feedback and refined UI for low-literacy staff
• Adjusted confidence thresholds based on real-world accuracy

Phase 4: Deployment (8 weeks)

• Phased rollout across main hospital and 8 village locations
• Ongoing training and support during first month
• Weekly calibration meetings to review AI accuracy
• Infrastructure improvements (solar backup for tablets, portable WiFi hotspots)

Technical Architecture

Frontend:

• React Native mobile app for Android tablets
• Offline-first architecture with background sync
• Simple, large-button interface for low-tech-literacy staff
• Visual feedback for AI confidence levels

AI/ML Components:

• TensorFlow Lite face recognition (runs on device)
• Custom OCR models trained on Indian identity documents (Aadhaar, ration cards)
• Confidence scoring for every extracted field
• Regular model updates based on staff corrections

Backend:

• Python FastAPI for high-performance API
• PostgreSQL with encryption at rest
• Redis for offline queue management
• FHIR-compatible data format for future interoperability

Infrastructure:

• Hosted on Indian cloud provider (Tata Communications)
• Edge processing on tablets for offline capability
• 4G dongles with local SIM cards for connectivity
• Solar-charged power banks for mobile camps
• Data residency in India for compliance

Security & Compliance:

• End-to-end encryption
• Audit logging for all data access
• Regular security assessments
• Staff access controls and training

Results

Operational Impact (12-month evaluation)

• About 2,000 patients registered across 8 villages
• 40% reduction in average registration time (from 18 min to 11 min)
• 95% data accuracy (up from 78%)
• 65% reduction in duplicate registrations
• Face recognition accuracy: 87% for suggesting correct matches (staff made final decisions)
• OCR accuracy: 89% for high-confidence fields, 94% after staff verification

Patient Experience

• Faster check-in for returning patients through face recognition
• Multi-language support removed communication barriers
• Reduced frustration with duplicate record issues
• Better continuity of care through unified patient records

Challenges Encountered

Technical:

• Power cuts: Resolved with solar power banks and offline-first design
• Internet connectivity: Offline queue holds up to 50 registrations for later sync
• Image quality: Staff training on proper lighting and document positioning improved capture quality
• False face matches: Initial threshold too low (adjusted from 75% to 82% similarity)

Human Factors:

• Initial staff resistance: Addressed through training emphasizing AI as assistant, not replacement
• Over-reliance on AI: Some staff accepting suggestions without verification (required retraining)
• Cultural barriers: Female health workers essential for women-only health camps

Human-in-the-Loop Design

Why Full Automation Wasn't Appropriate

Medical-Legal Responsibility:

• Patient identity verification carries legal implications
• Misidentification could lead to treatment errors
• Staff must be accountable for registration accuracy

AI Limitations Acknowledged:

• Face recognition accuracy varies with lighting, age changes, image quality
• OCR struggles with worn or damaged documents
• Rural Indian identity documents have high variability

Trust Building:

• Staff trust system more when they maintain control
• Patients feel more respected with human interaction
• Gradual adoption easier than complete workflow disruption

Staff Empowerment

AI as Assistant, Not Replacement:

• System highlights fields for attention but staff makes decisions
• Staff can override any AI suggestion
• Training emphasized critical thinking about AI suggestions

Feedback Loop:

• Staff corrections used to retrain models quarterly
• Registration clerks recognized as data quality experts
• Monthly review meetings to discuss AI performance

Technology Highlights

Face Recognition Implementation

# Face matching with human verification
class AssistedFaceRecognition:
    def __init__(self):
        self.model = load_face_model()
        self.similarity_threshold = 0.82  # Adjusted based on real-world testing

    async def suggest_matches(self, new_face_image):
        # Extract face encoding
        encoding = self.model.encode(new_face_image)

        # Find similar faces in database
        matches = self.search_similar(encoding)

        # Return suggestions for STAFF REVIEW
        suggestions = []
        for match in matches:
            if match.similarity > self.similarity_threshold:
                suggestions.append({
                    'patient_id': match.id,
                    'patient_name': match.name,
                    'similarity': match.similarity,
                    'last_visit': match.last_visit,
                    'thumbnail': match.thumbnail,
                    'requires_staff_decision': True  # Always
                })

        return suggestions  # Staff chooses correct match or "new patient"

OCR Confidence Scoring

class ConfidenceBasedOCR:
    def extract_document_fields(self, document_image):
        results = self.ocr_engine.process(document_image)

        fields = {}
        for field_name, extraction in results.items():
            fields[field_name] = {
                'value': extraction.text,
                'confidence': extraction.confidence,
                'requires_review': extraction.confidence < 0.85,
                'bounding_box': extraction.location  # Highlight in UI
            }

        return fields  # UI highlights low-confidence fields

FHIR Integration (Future-Ready)

• Patient resource structure follows FHIR R4
• Designed for future integration with national health systems
• Enables data sharing with district hospitals when patient referred

Lessons Learned

1. Human-in-the-Loop is Essential: Full automation inappropriate for patient identity verification
2. Offline Capability Critical: Internet unreliable in rural settings; offline-first architecture essential
3. Cultural Context Matters: Female health workers necessary for women's health camps
4. Staff Training is Ongoing: Initial training insufficient; continuous education required
5. Realistic Expectations: Set achievable goals (87% AI accuracy) rather than overpromising
6. Infrastructure Investment: Solar power, portable WiFi essential for rural deployment
7. Incremental Rollout: Phased deployment allowed adjustments based on real-world learning

Community Impact

Underserved Populations Served:

• 8 villages across rural Tamil Nadu
• Primary beneficiaries: Women, elderly, low-literacy patients
• Mobile health camps reach patients who cannot travel to main hospital

Improved Healthcare Access:

• Faster registration means more time with healthcare providers
• Unified records improve continuity of care
• Reduced duplicate records prevent medication errors

Sustainability:

• Local staff trained to maintain system
• Solar power reduces dependence on unreliable grid
• Cloud costs kept low through Indian hosting

Future Enhancements

• Fingerprint integration for patients without identity documents
• SMS notifications in Tamil for appointment reminders
• Integration with government health portal (Ayushman Bharat)
• Expansion to 15 additional villages (funding permitting)
• Teleconsultation integration for specialist referrals

---

This project demonstrates using AI responsibly in healthcare settings where human oversight is essential. The system augments staff capabilities while maintaining human accountability for patient care decisions.

Ready to implement human-centered AI for your rural healthcare organization? Contact us to discuss your needs.