Clinical intelligence for the AI era requires both expert and machine

The importance of AI not answering a question

A key differentiator when using generative clinical tools is whether the AI tool can recognize when it doesn’t have an answer to provide instead of hallucinating a response.

“In medical school, a key thing you’re taught is to learn when to say, ‘I don’t know,’” says Dr. Eckler. “Learning to be a surgeon includes learning when not to operate. It’s important that when we come to these systems for guidance that we get the response that is best for the situation, not the response the system thinks we want.”

To support the tool providing a non-answer, it’s critical to build safeguards, rubrics, and standards. With responses, traceable sources are critical for assessing whether it’s verifiable and defensible.

“This technology is transforming how we retrieve information. But synthesis is just as important. As patients, we aren’t immune to harm that can come from an AI tool trying to guess at an answer. These can impact real humans with real lives every day.”

Expanding the UpToDate editorial process for the AI era

With the release of UpToDate® Expert AI—evidence-based GenAI decision support—the trusted UpToDate Editorial Process has evolved. In addition to clinical experts informing the Editorial Process, they now also help inform the generated AI responses.

Together, this is the UpToDate Clinical Intelligence Model (CIM).

The clinician-authored editorial process is the continuous heart of the model. As new clinical studies emerge, our 7,600 clinical experts review them for application to the wider clinical context and grade the strength of the care recommendation. The proprietary UpToDate content is revised, peer-reviewed for consensus, and published to the CDS platform—establishing a unique foundation to create a purpose-built AI solution the industry can trust.

On top of that unique and valuable clinical intelligence foundation, our team of practicing experts have encoded their clinical experience into how UpToDate Expert AI thinks, and also continuously evaluate the clinical performance. That critical enablement layer involves:

• Clinical questions are asked within UpToDate Expert AI
• Clinical expertise is encoded, and guides question comprehension and relevance
• The AI responses are accountable and transparent with direct sources to UpToDate content and with clinician input
• Expert-developed rubrics and guardrails guide the response framework and language
• AI responses are refined with clinician input and feedback

“Our established editorial process has over 30 years of vetted value—we’re not throwing that out,” says Dr. Eckler. “In fact, it’s more important than ever. We’ve got more information than ever coming our way. We need to make sure the machines we trust are generating responses grounded in a clinical model, not a language probabilistic model.”

Dr. Eckler says the opportunity for AI use in patient care has incredible long-term potential, but the gravity and consequence of decision-making—as well as keeping the expert-in-the-loop—must remain central to the conversation.

“AI is a phenomenal tool, and we are still learning how to best use it. But this tool doesn't negate the history of clinical thinking by humans in what we've been doing thus far. It's going to take us forward if we use it the right way. It's incredibly impressive, but it's not replacing the history of clinical thinking.”

Download our overview on the Clinical Intelligence Model for more details on how we build, verify, and maintain the trusted content of UpToDate in the AI era. Explore how UpToDate Expert AI can extend evidence-based clinical information across your organization.