From RAG To Rich Context - Bringing Personalized AI to Patients with MCP

It feels like just yesterday, we were all excited about Retrieval Augmented Generation (RAG) -and for good reason! RAG was a monumental leap, finally giving large language models like ChatGPT the ability to tap into external knowledge and provide context-aware responses. It promised a new era of AI in healthcare, where models could assist with information grounded in real data.

Now, just a few months later, we're not just talking about RAG anymore. We're immersed in discussions around the Model Context Protocol (MCP), a paradigm shift that takes contextual understanding to an entirely new level - think 100x richer, more precise, and more integrated.

If you haven't followed the conversations, you might wonder, "What exactly is MCP, and why is everyone so excited?" MCP offers a standardized, robust way for AI models to retrieve context when needed. It's about creating a truly "rich context" that empowers AI to understand the nuances of healthcare data in a way we didn't even dream of. At the end of this newsletter, I've shared some resources we have been working on for those who want to dive deeper.

MCP integration with EHRs

In my last post, I explored the various options for leveraging MCP in clinical workflows with EHRs (a summary image is attached below). I've emphasized the critical need to design standardized MCP servers that can be leveraged across diverse EHR (and other healthcare) systems. This approach would allow for a model similar to SMART on FHIR to flourish, enabling plug-and-play interoperability for MCP servers.

When I refer to MCP servers, I'm specifically advocating for remote MCP servers utilizing protocols like HTTP with SSE (not Stdio). IMO, this is the only way to achieve comprehensive, vendor-agnostic enterprise-level interoperability.

The feedback and interactions from those discussions have been invaluable, and I'm deeply grateful for the engagement. However, that post also sparked some new, fascinating questions: While much of the excitement around advanced AI in healthcare has focused on provider-facing services, what can MCP do for patients when considering patient-facing services and tools built around their unique context?

What does MCP for the patient truly mean?

Patients have had to navigate a fragmented healthcare system for too long, often repeating their stories and medical histories to every new provider and still not getting satisfactory, detailed answers. Patient access to general AI tools like ChatGPT has been an enormous step forward in enabling them to find information. While concerns about safety and inaccurate advice persist, it's still a preferable alternative to Dr. Google for many. As great as ChatGPT (and other LLMs) have been for patients, it hasn't completely solved the problem of personalized, integrated care. Patients are still manually typing or copy-pasting pieces of their information and trying to get answers.

Envision AI intelligently knowing when and how to retrieve relevant information from a longitudinal patient record based on the real-time context of the conversation. This is what MCP can enable.

Imagine if AI seamlessly integrated and understood a patient's entire healthcare journey without them having to type it all out. No, we're not talking about patients manually uploading every document. Instead, envision AI intelligently knowing when and how to retrieve relevant information from a longitudinal patient record based on the real-time context of the conversation. This is what MCP can enable. We can build a new generation of patient-facing AI services by leveraging MCP. Again, I am referring to remote MCP servers. We can't expect all patients to be developers and configure local MCP servers. These aren't just chatbots; they're intelligent companions that understand the unique context of each patient's health. Think about:

• Proactive tools that learn from their data and offer personalized insights.
• Symptom checkers that can access your medical history for more accurate guidance.
• Appointment scheduling and follow-ups that are truly tailored to your needs.
• Personalized health education that evolves with your condition and questions.

This is a profound shift from patients using AI to benefiting from AI services deeply interwoven with their data.

It's about moving beyond general information to truly intelligent, personalized support that understands the full "context" of a patient's health journey.

This vision can be achieved with a variation of option four, outlined in the image above. Rather than bulk EHR exports, patients or caregivers aggregate data via consumer-mediated exchange from diverse sources like providers, payers, and devices in an online personal repository. The rest of the process can be the same, with the patient interacting directly with this rich data instead of the clinician.

These repositories can then be connected through MCP servers, highly specialized for different patient profiles or even condition-specific communities. This decoupling of research and contextual data, made available directly to patients, opens up amazing opportunities for highly personalized support. While robust standards and security will be paramount, the potential for empowering specific patient communities with tailored, relevant information is genuinely transformative.

This paradigm powerfully underscores my proposal that MCP servers should be developed independently of EHRs or other healthcare systems. Many of these future MCP servers will be versatile enough for use cases that can safely benefit both clinicians and patients, each receiving the appropriate context.

I have to admit, when this idea truly clicked from the post comments, the sheer potential of remote MCP for patients, I was so energized that even the Memorial Day weekend couldn't stop the internal brainstorming. I pulled the team together for a session to dive into what we could support and build into the core framework we have been working on around MCP! Some fascinating discussions to share in future posts.

Join the Rich Context Revolution!

We at Darena Health have been working on a framework to enable the development of plug-and-play MCP servers with any healthcare system (e.g., EHR). In the coming week, we'll share what we've built and launch a community project to accelerate the development of various use case-based MCP servers. We're getting ready to open-source this initiative and can't wait to start working with everyone.

Connect and Explore More

If you're as excited as we are about these possibilities, especially for patient empowerment, we'd love to discuss collaboration! Please contact us or subscribe to our newsletter to stay connected.

We recently shared an early concept to give you a glimpse into our vision. In the video below, I explain MCP and the problem it's designed to solve, and I explore real examples of an external MCP working with an FHIR server.

Below are some of my recent posts that delve into various aspects of this transformative protocol.

5 Key Comparisons to MCP

MCP: Making LLMs SMARTer for Healthcare

What exactly is a Tool in MCP?

4 Ways to Integrate Copilots with EHRs

Intersection of SMART on FHIR and MCP