Clinical Documentation AI: 55 Minutes Saved Per Physician Daily
An ambient AI documentation agent using Whisper speech-to-text and Claude for clinical note generation that integrates directly with Epic EHR via FHIR R4 — restoring physician time and reducing burnout.
The Challenge
The Challenge: Physicians Spending More Time Documenting Than Treating
A 45-physician multi-specialty group practice was facing a physician burnout crisis with documentation as the primary driver. Survey data showed physicians spending an average of 2.1 hours per day on clinical documentation — chart completion, progress notes, referral letters, and after-visit summaries — after their clinic schedule ended. 78% reported that documentation burden was the top factor affecting their job satisfaction. Two physicians had resigned citing administrative overload. The practice was using Epic as their EHR but had not adopted any AI documentation tools, viewing the available ambient scribing solutions as either prohibitively expensive (subscription costs of $500-$1,000 per physician per month for commercial solutions), insufficiently customisable to their specialty-specific documentation requirements, or raising HIPAA concerns with third-party audio processing. The requirements were clear: reduce documentation time, maintain clinical note quality, integrate natively with Epic, and keep patient audio within their own HIPAA-compliant infrastructure.
The Solution
Solution: Self-Hosted Ambient AI Documentation with Epic FHIR Integration
I built a self-hosted ambient clinical documentation system using OpenAI Whisper for speech transcription and Claude 3.5 Sonnet for clinical note generation — deployed within the practice's Azure tenant so patient audio never leaves their HIPAA Business Associate Agreement boundary.
Ambient Audio Capture & Transcription
A lightweight desktop application (Windows/Mac) that physicians activate at the start of each patient encounter. Audio is captured locally, processed through Whisper large-v3 running on an Azure GPU instance within the practice's subscription, and returned as a timestamped transcript. Speaker diarisation separates physician and patient speech. The application also accepts dictated notes via a push-to-talk mode for after-encounter addenda.
Specialty-Specific Note Generation
Claude 3.5 Sonnet processes the encounter transcript with specialty-specific system prompts for each of the practice's seven specialties (family medicine, internal medicine, cardiology, orthopaedics, dermatology, endocrinology, and gastroenterology). Each specialty has custom prompt templates that match their documentation requirements, preferred note structure, and medical terminology patterns. Claude generates a structured SOAP note with HPI, ROS, physical examination findings, assessment, and plan — including ICD-10 diagnosis code suggestions and CPT code recommendations for the encounter.
Epic FHIR Integration
Built a .NET 8 integration service that pushes generated notes directly to Epic via the Epic FHIR R4 API, creating ClinicalNote resources that appear in the physician's Epic In Basket for review and sign-off. The integration also reads the patient's existing FHIR data (active problem list, current medications, recent lab results) to provide Claude with patient context that improves note accuracy. Signed notes update the patient record automatically.
Quality & Compliance Layer
Implemented a quality scoring system that flags generated notes for physician attention where confidence is low (ambiguous audio, medical terminology conflicts, or inconsistency with existing patient data). All audio files are encrypted at rest, automatically purged after 24 hours, and the system maintains a complete HIPAA audit log of data access and note generation events.
Technology Stack
Tools & Technologies Used
“I go home on time now. That sounds small but after years of spending two hours every evening finishing charts, it has completely changed my relationship with my work. The notes are better than what I was writing manually because I am focused on the patient during the encounter, not on what I am going to have to document afterward.”
Related Services
Services Used in This Project
Have a similar project?
Let's discuss your requirements. Book a free 30-minute consultation.
Book Free Consultation