Healthcare AI News 3/12/25

News

OpenAI introduces a developer platform for building AI agents that includes tools to perform web and file searches and to perform web-based tasks similar to its Operator browser.

NHS England is deploying an AI tool that can predict a patient’s risk of falling with 97% accuracy. The software, which was developed by Cera, is also being used to predict deterioration in home care patients.

A study finds that patients slightly preferred AI-generated responses to their portal questions over human-written ones, but reported lower satisfaction when told that the response came from AI. The authors conclude that patients should be told that AI was used since it didn’t reduce satisfaction significantly. They also polled patients on their preferred wording of the disclosure, with the winner being, “This message was written by Dr T. with the support of automated tools.”

Business

Memorial Sloan Kettering Cancer Center completes a pilot of Abridge’s AI ambient documentation and plans a broad rollout over the next two years.

AI drug discovery company Insilico Medicine deploys a “bipedal humanoid” to train AI systems on the tasks performed by laboratory scientists. They are also using the robot, called “Supervisor,” to assist with lab tours, telepresence, and lab supervision.

Research

Researchers find that LLMs show promise in reducing pediatric medication dosage errors. A medication ChatGPT and Claude were more accurate and faster than pediatric and neonatal nurses, while Llama performed poorly due to an apparent weakness in its calculation logic. The authors recommend evaluating specific LLMs rather than treating all of them as equally capable.

LLMs exhibit “anxiety” when processing emotional mental health topics like interpersonal violence and accidents. Researchers found that mindfulness-related prompts could help regulate the model’s responses, similar to how human therapists manage their emotional reactions while maintaining empathy.

Stanford researchers use AI to identify a naturally occurring prohormone that is as effective as Ozempic in weight loss without the side effects of nausea, constipation, and loss of muscle mass.

Other

Patients are using LLMs to analyze their hospital bills for charges that exceed state and national averages. New startup OpenHand is offering similar analysis, after which the company negotiates with providers to lower the bill.

TikTok users report that AI-generated deepfake doctors are spreading medical advice on topics like surgery, diet, and cosmetic procedures. Some use the Captions app to create and edit AI videos that can be easily replicated with different messages, which is how the users noticed the fakes.

Epic and Agentic AI

A reader asked for my take on Epic’s plans for agentic AI. I have no inside knowledge, so this is pure spitballing.

Some background. Agentic AI acts independently to achieve goals without human oversight while responding dynamically to its environment. Think self-driving cars. It renders robotic process automation (RPA) obsolete, as RPA relied on rigid rules and predefined inputs. It’s hard to believe it’s been just four years since Olive was health tech’s hottest startup.

Non-agentic AI, by contrast, requires human direction. Chatbots are an example. They answer questions and retrieve information but don’t take external actions like scheduling appointments. In between are limited function, app-specific copilots that assist users without initiating decisions.

The business case for agentic AI is workflow automation, reduced labor costs, real-time monitoring (cybersecurity, throughput, resource allocation), and rapid feature deployment. Instead of modifying core systems via traditional coding, testing, and releases, AI can introduce new functionality faster and allow customization at the client level. It also streamlines integrations with external systems. All of this is theoretical, of course, and is heavily dependent on the vendor and user organization.

Epic has already embedded non-agentic AI across its platform, with use cases like drafting patient replies, simplifying documents, automating prior authorizations, and enabling voice control. These are quickly becoming table stakes with AI’s ubiquity.

Agentic AI is the logical next step, and Epic seems to be out front, although Oracle Health’s plans aren’t quite clear yet either. Early implementations will likely focus on low-risk back-office tasks, then expand into clinical support, population outreach, and automated reminders. Unlike third-party AI vendors that rely on brittle workarounds like screen scraping, Epic can integrate AI natively and provide scalability and stability.

AI’s role in clinical decision support is gaining acceptance, as long as a human remains in the loop as FDA requires to avoid inviting regulation as a medical device. Future AI applications could preassemble patient histories, flag care gaps, match patients to clinical trials, and pull relevant literature. AI could also be used to personalize the patient’s treatment and communication.

Few vendors have the resources to develop and support AI agents that have unknown ROI. Reputational risks from AI errors and regulatory scrutiny will be a deterrent for some companies. Another possible barrier is the willingness of a developer-focused software company to allow an AI agent to take over software flow but still support normal user interaction.

Epic benefits from its homogeneous customer base and a track record of incremental software development. It doesn’t need to chase AI-jazzed investors, so it can roll out tools when it’s ready in an Minimum Viable Product-type approach. 

Epic also has advantages such as its Cosmos data repository, the ability to integrate deeply with its existing products, and the market power to influence what partners and competitors do.

I would expect Epic to deploy both agentic and non-agentic AI initially to reduce clinician burden and surface relevant insights within workflows. It will probably have another group working on reducing the health system labor that is needed to basically push (electronic) paper that someone outside the health system requires. It will eventually use AI to adapt its underlying software to user preferences. It will probably tread lightly at first with clinical functions, making sure to allow opt-outs and human overrides when the AI’s confidence is low.

On the big-picture operational side, Epic will position itself as offering an intelligent, proactive platform for hospital management, which people have been talking about for years. That will be a significant development assuming that early adopters show measurable improvement in moving from “tools” to “systems.”

Success depends on Epic’s ability to build new expertise in AI and determine the level of cloud dependency its customers will accept. It’s likely already working with an early adopter cohort, though we won’t hear much outside of UGM presentations. By August, we should have a clearer picture of its direction. Anything in the meantime is speculation, which I wouldn’t have offered if the reader hadn’t asked. Your thoughts are welcome.

Contacts

Mr. H, Lorre, Jenn, Dr. Jayne.
Get HIStalk updates.
Send news or rumors.
Follow on X, Bluesky, and LinkedIn.
Contact us.