Revolutionizing Healthcare Documentation with AI Medical Scribe

Revolutionizing Healthcare Documentation with AI Medical Scribe

Table of Contents:

1. Introduction
2. The Problem of Documentation in the Medical Field
3. The Solution: AI Medical Scribe
4. The Journey of Circle Medical
5. The Challenges Faced Along the Way 5.1 Transcription Capture 5.1.1 Integrating with Video Call Vendor API 5.1.2 Hiring a Transcription Bot 5.1.3 Building a Desktop App or Web App 5.2 Transcript Identification 5.2.1 Ensuring Privacy and Compliance 5.2.2 Available Identification Solutions 5.3 Generating the Clinical Note 5.3.1 Using GPT-4 to Generate Clinical Notes 5.3.2 Addressing Hallucination Issues 5.3.3 Giving Additional Context for Better Results 5.3.4 Incorporating Clinical Guidelines 5.4 Evaluating and Iterating to Improve Results 5.4.1 Human Feedback 5.4.2 Comparing to Ground Truth 5.4.3 Using Grading Criteria Prompt
6. Tools for AI Development in Ruby
7. Conclusion
8. Job Opportunities at Circle Medical

Introduction

In the ever-evolving field of Healthcare, continuous efforts are being made to improve the quality, cost, and accessibility of primary care. One of the challenges faced by healthcare providers is the extensive time spent on documentation, particularly in the form of clinical notes. To address this issue, Circle Medical, a San Francisco-based company, embarked on a journey to develop an AI medical scribe. This article explores the problems encountered, the solutions devised, and the journey of Circle Medical in implementing this AI-powered solution.

The Problem of Documentation in the Medical Field

Traditionally, the quality, cost, and accessibility of healthcare were believed to be an indivisible trade-off. However, Circle Medical challenged this Notion and aimed to leverage technology to break the so-called "Iron Triangle" of healthcare. One of the significant pain points in healthcare documentation is the time-consuming process of physicians documenting their visits and sending orders at the end of the day. This administrative burden on physicians leads to burnout and affects the overall quality of care. Furthermore, the documentation serves various purposes, including patient education, legal compliance, continuity of care, and billing. As Circle Medical grew, maintaining the quality and consistency of documentation for a multitude of providers became a challenge. The need for a solution that could address these issues became evident.

The Solution: AI Medical Scribe

Circle Medical set out to develop an AI medical scribe, an innovative solution that would listen to patient-provider conversations and generate clinical notes automatically. While hiring human scribes was an option, it proved to be expensive, disruptive, and imperfect. Leveraging the advancements in Natural Language Processing (NLP) and the availability of powerful language models such as GPT-4, Circle Medical aimed to create a system that could accurately and efficiently capture and generate clinical notes. By integrating this AI technology into their Electronic Medical Records (EMR) system, Circle Medical aimed to provide a seamless experience for physicians while maintaining the quality and consistency of documentation.

The Journey of Circle Medical

Circle Medical began its journey in 2015 in San Francisco, growing steadily over the years. With a team of 25 engineers and 350 doctors, Circle Medical focused on making primary care accessible to everyone by leveraging technology. They developed a Ruby on Rails stack with React Next.js and built native apps for patients. During the pandemic, the acceptance of telemedicine skyrocketed, and Circle Medical was at the forefront, seeing over 50,000 virtual patients per month. However, the burden of documentation remained a challenge, leading them to embark on their AI medical scribe journey.

The Challenges Faced Along the Way

Transcription Capture

To develop an AI medical scribe, the first step was to capture the patient-provider conversations accurately. Circle Medical explored various options for capturing the transcript, each with its pros and cons. Integrating with existing video call vendor APIs, hiring transcription bots, building desktop or web apps, and using third-party transcription services were all considered. Ultimately, Circle Medical decided to integrate with their video call vendor, Zoom, to capture the transcripts directly. This approach provided a seamless experience for patients and physicians, albeit with variable transcription quality and limited real-time transcript availability.

Transcript Identification

With the captured transcripts, the challenge of privacy and compliance emerged. To ensure patient data confidentiality and comply with HIPAA regulations, Circle Medical needed to identify sensitive information in the transcripts before using them for generating clinical notes. Various solutions ranging from using AI identification services to building in-house identification models were explored. Ultimately, Circle Medical opted for an AI identification service called Private AI, which provided fine-grained control over what information needed to be identified to ensure privacy.

Generating the Clinical Note

The core of the AI medical scribe system was the generation of clinical notes based on the transcripts. Circle Medical used GPT-4, a powerful language model, to generate the clinical notes. However, they faced challenges such as hallucination (the model making things up), inconsistent output, and the lack of adherence to clinical guidelines. To address these issues, Circle Medical explicitly instructed the model not to make assumptions or recommendations not Mentioned in the transcript. They also discovered that providing additional context, such as patient demographics, problem lists, and active medication lists, significantly improved the accuracy and consistency of the generated clinical notes.

Evaluating and Iterating to Improve Results

Iterating and improving the AI medical scribe system required effective evaluation techniques. Circle Medical employed three strategies: human feedback, comparison to ground truth, and grading criteria prompts. Human feedback from physicians who reviewed the generated clinical notes provided valuable insights into the quality and accuracy of the notes. Comparison to ground truth allowed Circle Medical to measure the similarity between the generated notes and manually created high-quality clinical notes. Finally, the grading criteria prompts enabled Circle Medical to assign scores based on specific evaluation criteria, facilitating a more objective assessment of the generated clinical notes.

Tools for AI Development in Ruby

Throughout their AI medical scribe journey, Circle Medical leveraged various tools in their Ruby-based development stack. Jupyter Notebook, a popular tool among data scientists, proved useful for running Ruby code and integrating with the Rails app. They also relied on the Daru gem, a Ruby equivalent of the pandas library in Python, for data analysis. Prompt engineering played a crucial role, and Circle Medical found resources like OpenAI's Prompt Engineering Guide helpful in optimizing their prompts for better results.

Conclusion

Circle Medical's journey to develop an AI medical scribe highlights the challenges faced in healthcare documentation and the potential of AI to address these challenges. By capturing accurate transcripts, identifying sensitive information, and generating clinical notes with the help of GPT-4, Circle Medical aims to revolutionize the documentation process in primary care. Continuous evaluation and iteration are essential to ensure the system's accuracy and adherence to clinical guidelines. The use of Ruby-based tools and prompt engineering techniques has played a significant role in Circle Medical's success thus far.

Job Opportunities at Circle Medical

Circle Medical is actively seeking talented individuals to join their engineering team. As they plan to double the team's size in 2024, various positions will be available. Interested individuals can visit the Circle Medical website or reach out directly to JS at Circle Medical for more information about the exciting job opportunities within the company.

🔎 Resources:

• Circle Medical website: www.circlemedical.com
• Private AI: www.private.ai
• Recall AI: www.recall.ai
• OpenAI's Prompt Engineering Guide: OpenAI Blog - Prompt Engineering