Back to Portfolio
AI Product Development · Voice AI · HIPAA-Compliant RAG · EHR Integration

ClarityNote — AI Clinical Documentation Assistant for NovaCare Health Network

An ambient AI system that listens to physician-patient consultations, generates structured SOAP notes in real time, and pushes them directly into Epic — saving NovaCare's 180 physicians 3.1 hours per day across 12 Dubai clinics, with full DHA regulatory compliance.

Whisper Large v3 ASR GPT-4o (Azure Private) Epic FHIR API WebSocket Streaming DHA Compliance React Native (iPad)
3.1 hrs
Saved Per Physician Per Day
94%
Note Accuracy vs Manual (validated)
12 Clinics
Deployed Across NovaCare
14 Weeks
Concept to Full Deployment
Client
NovaCare Health Network
Industry
Healthcare / Clinical Operations
Location
Dubai, UAE
Duration
14 Weeks · Q2–Q3 2024
Project Overview

About This Project

NovaCare's physicians were spending 3–4 hours per day on clinical documentation — typing notes into their EHR (Epic) after each patient consultation. This administrative burden was causing physician burnout, reducing patient-facing time, and creating a backlog of incomplete notes that delayed billing cycles by an average of 4.2 days.

We built ClarityNote: an ambient AI clinical documentation assistant that listens to the physician-patient conversation (with explicit patient consent), generates a structured SOAP note in real time, and pushes it directly into Epic for physician review and one-tap approval. Deployed across all 12 NovaCare clinics in 14 weeks, compliant with Dubai Health Authority (DHA) regulations and UK GDPR equivalents for international patients.

Whisper Large v3
GPT-4o (Azure Private)
FastAPI + WebSockets
React Native (iPad)
Epic FHIR API
AES-256 Encryption
Azure Private VPC
Redis Streams
Custom NLP Post-Processor
UAE DHA Compliance
3.1 hrs
Saved per physician per day — allowing an average of 4 additional patient consultations daily
94%
Clinical note accuracy against manual documentation — validated by NovaCare's Chief Medical Officer across 840 notes
47 sec
Average time from end of consultation to complete SOAP note ready for physician approval
12
NovaCare clinics fully deployed — 180 physicians across GP, Cardiology, Paediatrics, and Emergency departments
The Problem

Challenges We Solved

Medical Privacy Compliance in UAE

Healthcare AI in the UAE must comply with Dubai Health Authority (DHA) regulations — and for international patients, GDPR equivalents. Audio recordings of consultations require explicit consent flows, AES-256 encrypted storage, and strict retention policies with a 24-hour post-approval deletion requirement.

Clinical Accuracy Non-Negotiable

A misheard medication dosage or missed diagnosis detail in an AI-generated note could cause patient harm. The NovaCare CMO required a formal clinical validation study demonstrating accuracy comparable to manual note-taking before any physician deployment.

Accented Speech + Medical Terminology

NovaCare physicians include Arabic, Indian, Filipino, and British English speakers discussing highly technical medical terminology. Standard Whisper had 12% error rates on medical terms including "metformin", "cholecystitis", and "nasopharyngeal carcinoma" in initial testing.

Real-Time Processing Requirement

Physicians wanted to see a structured note forming on their iPad screen during the consultation — not after. This required streaming ASR with real-time LLM processing with sub-3-second latency from speech to on-screen text, under variable clinic WiFi conditions.

Epic EHR Integration Complexity

Epic's FHIR API required SMART on FHIR OAuth 2.0 certification, specific note templates per department (GP, Cardiology, Paediatrics, Orthopaedics, Dermatology, Emergency), and automatic ICD-10 code suggestions — with Epic's approval process adding 3 weeks to the integration timeline.

Physician Trust and Adoption

Healthcare workers are deeply sceptical of AI tools following multiple high-profile failures. Any note that required heavy editing would be immediately abandoned. The system had to generate notes that felt like the physician wrote them — not "AI-sounding" summaries.

Our Approach

How We Solved It

DHA-Compliant Consent + Encryption Architecture

Built a pre-consultation consent flow on a patient-facing iPad screen in Arabic and English. Audio is AES-256 encrypted at the microphone level, transmitted over TLS 1.3, processed exclusively in Azure UAE North region, and auto-deleted within 24 hours post-approval. Full audit trail maintained for 7 years per DHA requirements — passed DHA compliance review with zero audit findings.

Clinical Accuracy Validation Protocol

Ran a formal 6-week clinical validation pilot with 3 NovaCare physicians reviewing 840 AI-generated notes against their own contemporaneous manual notes. Iterated prompt engineering and post-processing rules through 12 revision cycles until the NovaCare Chief Medical Officer formally validated 94% accuracy — documented in a clinical study report submitted to DHA.

Medical ASR Fine-Tuning on 2,400 Transcripts

Fine-tuned Whisper Large v3 on a corpus of 2,400 consented, de-identified medical consultation transcripts from NovaCare's archive. Added a medical terminology correction layer with a 45,000-term medical dictionary for post-ASR normalisation. Error rate on medical terms dropped from 12% to 1.8% — reducing physician correction time to under 90 seconds per note.

WebSocket Real-Time Streaming Pipeline

Built a real-time pipeline: iPad microphone → WebSocket → ASR chunk processing in 100ms sliding windows → incremental SOAP note section generation → streaming display to iPad. Physicians see the note forming section by section during the consultation — Subjective, Objective, Assessment, and Plan populating in real time.

Epic SMART on FHIR Integration

Achieved Epic App Orchard certification with SMART on FHIR OAuth 2.0. Supports 6 department-specific note templates (GP, Cardiology, Paediatrics, Orthopaedics, Dermatology, Emergency) with automatic ICD-10 and CPT code suggestions pulled from a continuously updated medical coding database — one-tap approval pushes directly into the patient's Epic chart.

Per-Physician Voice Adaptation

Implemented per-physician voice profiling: after 20 approved notes, the system learns each doctor's preferred terminology, abbreviations, note structure, and level of clinical detail. Notes become increasingly personalised — physicians reported that notes "sounded like them" within 2 weeks of use, dramatically increasing approval rates.

Results

The Outcomes

Week 8 — Validation
94% Accuracy Validated by CMO

Clinical pilot with 12 physicians and 840 reviewed notes. The NovaCare Chief Medical Officer formally signed off 94% accuracy — above the 90% threshold required for deployment approval. DHA compliance team issued clearance with no audit findings.

Week 14 — Deployment
All 12 Clinics Live

Full deployment across all 12 NovaCare clinics. 180 physicians onboarded across GP, Cardiology, Paediatrics, Orthopaedics, Dermatology, and Emergency departments. Training time per physician: 45 minutes. Average note accuracy at full deployment: 94.2%.

Ongoing — Impact
3.1 Hours Saved · 28% More Patients

3.1 hours saved per physician per day. NovaCare operations reporting a 28% increase in patient consultations across the network — physicians seeing an average of 4 additional patients per day with the documentation burden removed.

3.1 Hours Back Per Physician Per Day.
Across 180 NovaCare physicians, ClarityNote recovers the equivalent of 23 full-time physician work-days every single day — deployed DHA-compliant in 14 weeks.
★★★★★
"I was deeply sceptical. I've seen five 'AI clinical tools' fail in the last three years. ClarityNote is different — it actually sounds like me, it integrates with Epic seamlessly, and our DHA compliance team signed off without a single audit finding. I now see 4 more patients per day. The ROI paid for the entire project in the first month."
PN
Dr. Priya Nambiar
Head of General Practice, NovaCare Health Network Dubai

Build Your AI Product with Digivance

Ready to turn your AI idea into a production-grade product? Let's talk.

Back to All Portfolio Case Studies