Apple and OpenEvidence will test whether health AI can help patients prepare for doctor visits
Apple’s U.S. study will examine whether people understand and trust AI answers, and what happens when they choose to add data from their own health records and devices.
Apple and OpenEvidence are starting a U.S. study to see whether a health chatbot can help people get ready for an appointment with their own doctor. Participants can ask health questions and choose whether to use personal data from Apple’s HealthKit when they do. The study will examine what makes people trust the answers, whether they understand them and how good the responses are, Apple health and fitness vice president Dr. Sumbul Desai said Sunday on the HLTH USA Main Stage in Las Vegas.
Desai announced the Apple Health AI Usage Study after moderator Liz Baker Plosser, founder of the Best Case Scenario newsletter, invited OpenEvidence CEO Daniel Nadler onstage. The conversation had started with prevention features Apple already offers through its devices. It then turned to a different problem. A device can record how someone sleeps or moves, but putting those readings next to a person’s questions about their health is a task that depends as much on how the person interprets the answer as on the data itself.
“You have to show your work,” Desai said. She said participants could bring everyday questions to the study app, try adding data of their choice and explore whether the conversation makes them better prepared when they see a doctor. Apple’s study announcement says the app will show citations to clinical literature alongside the answers. The company says eligible U.S. adults age 21 and older can enroll with a compatible iPhone, with room for up to 200,000 participants. That is an enrollment limit. Desai shared no study results onstage.
The option to share personal data gives the researchers a concrete question to investigate. A general question about sleep can be answered from medical literature. A question about a change in someone’s sleep has more context if that person decides to make several weeks of readings available. The study can examine whether that context makes an answer clearer or more relevant to the user. Desai also put the quality of the response on the research agenda, an important distinction when an answer that feels more personal could also be more persuasive.
Apple lists activity, sleep, resting heart rate and lab results as examples of the data participants may elect to share. The company says sharing those metrics and answering surveys are voluntary, and that study data will be excluded from AI model training. Apple also says the study app is designed to flag certain concerning questions and point participants toward a provider or other resources. That description sets out what the company intends the app to do. How people understand and use a response is part of what the research is designed to examine.
Nadler said the opportunity came from physicians’ feedback about how much they can cover in a brief visit. It matters, he said, that “patients come ready and prepared to the visit.” He recalled taking his father to appointments and hearing doctors ask whether his father had experienced symptoms such as lightheadedness. When he could not say, he would have to go home, pay closer attention and book a follow-up. Nadler said better access to information gathered between visits might make an appointment more productive. His example described the clinical interaction he hopes to improve.
The example points to a question a clinic could recognize even without an AI tool. What information does a patient need to collect before the appointment so the physician and patient can spend more of their time deciding what comes next? A wearable may help show the pattern of a measurable signal. A symptom such as lightheadedness still depends on what the person notices and tells the clinician. A study about visit preparation can look at whether an AI interaction helps someone identify and communicate the details that matter.
Nadler described the broader opportunity as making better use of “what’s already known.” He argued that medical discoveries can have limited reach if the people who could use them have trouble finding or understanding them at the point of care. For a physician using OpenEvidence, the task might be finding the applicable literature during a visit. For a patient in the Apple study, it may begin with understanding an answer well enough to ask a better question. The same clinical literature reaches two audiences with different training and different decisions to make.
Earlier in the discussion, Plosser asked how Apple converts measurements into something people can act on. Desai used Health Age, a feature of the redesigned Health app, as an example. Apple looked at evidence for the measures it uses and drew on data from about 200,000 research participants as it explored how to combine them, she said. Desai framed the resulting metric as a way to prompt healthier habits. That earlier research was about inputs for a different feature. The new study is about what happens when a person poses a question and receives an answer in a conversation. Desai also discussed using the iPhone camera for movement assessments, such as a sit-to-stand test. The person performing that assessment has an immediate task and a specific measurement to review. The research app opens up a broader interaction in which the user decides what to ask and what context to provide.
Desai said the aim was to “build trust through privacy and clinical rigor.” She pointed to the challenge of interpreting several signals together, including sleep, heart rate variability and lab data. A person looking at each of those measurements separately might have trouble knowing what to ask at the next appointment. The study’s design lets Apple and OpenEvidence explore what a conversational response adds when the person chooses to share that context, and how well the person understands the answer it produces.
Apple announced a way to study those interactions. The test for healthcare organizations is more specific than whether patients are willing to try a chatbot. It is whether the tool can help people bring clearer questions and a better account of what they have experienced into a visit, while leaving them in control of the data they share. Nadler’s example makes that outcome concrete. A better-prepared patient could spend less of a brief visit reconstructing the past few weeks and more of it talking through what to do next.