TLDR: A custom-built AI scribe application, integrated into Included Health’s EHR, uses Whisper for transcription and GPT-4o for generating SOAP notes and patient instructions. It significantly reduces clinician cognitive load and documentation burden in telehealth, with 94% and 97% of surveyed clinicians reporting positive impacts respectively. The system also employs a fine-tuned BART model to improve note conciseness, demonstrating AI’s potential to streamline medical documentation.
In the demanding world of healthcare, clinicians often face significant challenges, particularly the heavy burden of administrative tasks like documentation. This workload can lead to burnout, impacting both the well-being of healthcare providers and the quality of patient care. A recent research paper introduces an innovative solution: a custom-built ambient medical scribe application designed specifically for telehealth clinicians.
Addressing Clinician Burnout with AI
The core problem identified is the time-consuming and complex process of creating detailed SOAP (Subjective, Objective, Assessment, Plan) notes for every patient visit. Modern Electronic Health Record (EHR) systems, while essential, often add to this burden due to their intricate interfaces and frequent updates. This administrative load is a major contributor to clinician burnout.
To combat this, Included Health, a personalized all-in-one healthcare company, developed an AI-powered scribe application. This tool is seamlessly integrated into their existing EHR system, aiming to streamline documentation and free up clinicians to focus more on patient care.
How the AI Scribe Works
The application leverages advanced AI models to automate the note-taking process. It uses Whisper, an open-source transcription model, to accurately transcribe visit audio. Following transcription, a modular in-context learning pipeline, powered by GPT-4o, automatically generates comprehensive SOAP notes and patient instructions. This modular approach allows for parallel generation of different sections of the note, such as the History of Present Illness and Assessment and Plan, improving efficiency and coherence.
A unique aspect of this system is its post-processing step. The researchers found that while the AI-generated notes were high quality, clinicians often edited them for conciseness. To address this, they fine-tuned a BART model to automatically refine the notes, reducing their length while preserving the essential semantic meaning. This step helps to further reduce the need for manual edits by clinicians.
Impressive Results and Adoption
The application underwent rigorous testing and evaluation. On mock visit data, the notes generated by the AI scribe were found to exceed the quality of expert-written notes, as determined by an “LLM-as-a-judge” evaluation method. This indicates a high level of accuracy and completeness in the AI’s output.
Perhaps most importantly, the application has seen widespread adoption and positive feedback from clinicians in a real-world production environment. Over 540 clinicians at Included Health have used the application. A survey revealed that a remarkable 94% of clinicians reported reduced cognitive load during visits, and 97% reported less documentation burden when using the application. Within three months of its integration in November 2024, the scribe application was used to document nearly 70% of all virtual primary care (VPC) visits and 40% of urgent care (UC) visits.
The fine-tuned BART model successfully reduced the character length of the History of Present Illness section by 17% compared to the initial AI-generated notes, with only a minimal impact on semantic meaning, further validating the system’s ability to create concise yet comprehensive documentation.
Also Read:
- Unlocking Medical Data: A Systematic Review of Synthetic Clinical Text Generation
- Enhancing AI Clinical Note Quality Through Physician Feedback Checklists
The Future of AI in Healthcare Documentation
While the long-term impact of AI medical scribes is still being explored, these findings highlight the significant potential for AI systems to alleviate administrative burdens and support clinicians in delivering efficient, high-quality care. The paper emphasizes the importance of integrating such tools directly into existing clinical workflows to ensure seamless adoption and maximum benefit. For more detailed information, you can refer to the full research paper available at arXiv.org.
The development team also ensured compliance with HIPAA data privacy and security standards through Business Associate Agreements (BAA) with LLM vendors. Patient consent is required before using AI for documentation, and clinicians are coached to always review AI-generated notes for accuracy, especially to guard against potential “hallucinations” or inaccuracies.
