Your Patient Brought an AI Health Summary—What Should You Do Next?

Picture a hypothetical 58-year-old who sits down, apologizes, and slides three printed pages across the desk. A chatbot has reviewed eight weeks of her smartwatch data and her last lab results. Near the bottom of page two, in bold, it says her pattern is “consistent with early heart failure.”

She’s frightened. She also looks a little embarrassed, as if she expects to be scolded for asking a computer before asking you.

Most physicians have two reflexes in that moment, and neither one is good medicine. The first is to wave the pages away. The second is to let them set the agenda for the entire visit. Both hand the encounter’s control to the software, one by rejection and the other by deference.

I’d like to offer a third option: Treat the summary the way you would an outside hospital’s discharge paperwork—potentially useful, possibly wrong, and worth 10 disciplined minutes rather than 40 anxious ones.

Why This Visit Is Becoming Routine

Patient-generated data now reaches the exam room constantly. In a survey of 2,222 physicians across six countries, 97% said they’d reviewed wearable data in some capacity, yet no country reported integration rates above 6%.1 That gap is where the AI summary lives. The data reaches the exam room. The workflow for handling it doesn’t.

Patients are already doing this. What matters is what a careful clinician does with the document once it’s on the desk.

Step One: Find the Question Underneath

Before reading a single line of the summary, I ask one thing: What made you want to look into this?

The answer is often more useful than the document. The hypothetical patient might say she’s been getting winded on the stairs at work. Or she might say her sister was just diagnosed with a heart problem and she hasn’t slept well since. Those are two very different visits, and the printout can’t tell you which one you’re in.

This step also lowers the temperature. It tells the patient you’re taking her concern seriously, which is usually why she came in. It also keeps you anchored to a symptom or a worry instead of to whatever the chatbot decided to bold.

Step Two: Trace Every Claim to Its Source

AI summaries blend three kinds of content, often in the same paragraph. There’s raw data the patient’s device actually recorded. There’s the device maker’s own interpretation of that data, such as a proprietary “recovery” or “readiness” score. And then there’s the chatbot’s narrative layered on top.

Only the first category comes close to a measurement, and even that depends on the sensor, the fit, and the algorithm the manufacturer uses to turn raw signals into numbers. The second is a vendor’s algorithm, usually undisclosed. The third is a language model’s attempt to tell a coherent story, and coherence is exactly what these models are built to produce, whether the underlying signal supports it or not.

In practice, I’d ask the patient to open the app on her phone. What does the resting heart rate trend actually show, in beats per minute, over which dates? How many nights of sleep data are missing? Did she wear the watch consistently, including overnight, or did she take it off for a week while traveling? A few minutes of this usually tells you how much weight the summary can bear.

Step Three: Separate Observation from Inference

One habit I recommend to every clinician who handles outside information is almost embarrassingly simple. Rewrite the summary’s key claim in two columns, even if only in your head.

On the left, write what was observed: “Resting heart rate averaged 64 in July and 71 in August. Sleep duration fell by about an hour on weeknights.” 

On the right, write what was inferred: “This suggests early heart failure.”

Once the columns are split, the inference usually looks a lot weaker. A rising resting heart rate over a month has a long differential, and much of it is benign or situational—deconditioning, a new medication, poor sleep, alcohol, dehydration, anemia, thyroid disease, an intercurrent infection, or simple stress over a sister’s diagnosis. The chatbot picked one frightening explanation and wrote it with confidence. That’s the part to be set aside, not the numbers themselves.

The numbers may still deserve attention. That’s what makes this skill worth teaching.

Step Four: Look for What’s Missing

AI summaries are fluent about what they have and silent about what they lack. That silence is where the clinical risk hides.

A few gaps show up again and again. Symptoms are often absent entirely, because the patient never typed them in. Medications may be incomplete or out of date. The summary rarely knows about the week of travel, the new job, or the beta blocker she stopped on her own. And almost none of these documents report how much data is missing, which can matter as much as the data that’s there.

In emergency medicine, the most dangerous piece of information is often the one nobody mentioned. The same instinct applies here. Ask directly: What’s not on these pages that I should know?

Step Five: Decide What Earns Follow-Up

Now you’re on familiar ground. You’ve got a patient, a concern, a history, an exam, and a set of observations you’ve verified yourself. The AI summary has now become one input among several, which is where it belongs.

Sometimes the verified data does raise a flag. A sustained change in resting heart rate, alongside new exertional dyspnea and edema on exam, would warrant a proper cardiac workup regardless of what any chatbot said. Other times, the pattern dissolves once you know she stopped exercising after a knee injury.

Either way, the decision should rest on your clinical reasoning, not on the document’s confidence. If new or severe symptoms are in the picture, nothing in the printout should slow down urgent evaluation.

Document what you verified and what you concluded, in plain terms. A note that says “patient-provided AI summary reviewed, underlying heart rate trend confirmed, no exertional symptoms, plan as below” protects the patient, the next clinician, and you. It also creates a record that you weighed the document rather than ignored it.

Guard Your Own Judgment, Too

It’s easy to assume that physicians are immune to a persuasive machine. One recent study suggests otherwise.

In a randomized study published April 23, 2026, 44 physicians who had completed AI-literacy training worked through clinical vignettes with either error-free AI recommendations or recommendations containing deliberate errors. Those exposed to the erroneous recommendations performed worse on diagnostic reasoning.2 The cases were simulated, and no patients were involved, so the findings can’t be mapped directly onto clinic visits. Still, the direction of the effect should give every clinician pause.

A patient’s AI summary is a recommendation of sorts, and it arrives pre-loaded with a diagnosis. The step-by-step review above is partly a protection for the patient and partly a protection for you. Reading the raw observations before the narrative is the simplest defense I know against being quietly steered.

What to Say to the Patient

Tone matters as much as method. I’d avoid anything that sounds like a reprimand, because a patient who feels judged for using AI will simply stop telling you she’s using it. That’s a worse outcome than a flawed printout.

Something like this tends to work: “I’m glad you brought this. Some of what it found is worth looking at, and some of it goes further than the data can support. I’d like to sort out which is which with you.”

That sentence validates the effort, separates the signal from the story, and keeps you in the role of interpreter. It also sets up a better visit next time, because a patient who feels heard is far more likely to come back with the same question and a better set of data than to take her next printout somewhere else.

A Better Summary for the Next Visit

If patients are going to bring these documents anyway, I’d rather shape what they bring. I suggest a single page with five parts. First, the one question they most want answered. Second, the date range covered. Third, the two or three changes they noticed, in their own words. Fourth, any symptoms, with dates. Fifth, a current medication list.

That page takes a patient 10 minutes to prepare and saves you from reconstructing a timeline under pressure. It also nudges the patient to think before the visit, which may be the most valuable thing an AI tool can prompt.

One practical caution: Before copying a patient’s summary or raw data into any AI system of your own, check that the tool is approved by your organization for protected health information. The patient’s choice to use a consumer app doesn’t extend that choice to you.

The Encounter Still Belongs to Both of You

The patient who brings an AI summary is doing something physicians have wanted for decades. She’s engaged, curious, and trying to understand her own body. The document she hands you may be half right and half wrong, often in the same sentence.

What a physician can do is find the real question, verify what can be verified, name what’s missing, and decide what deserves care. Done well, that work turns a frightened patient holding a printout into a partner who knows exactly what to bring next time.

About the Author

Harvey Castro, MD, MBA, is a board-certified emergency physician, a five-time TEDx speaker, and the founder of DR GPT™. He has written over 45 books, including AI in Emergency Medicine, forthcoming from Wiley in October 2026. His work focuses on the practical, responsible use of artificial intelligence in healthcare.

www.HarveyCastroMD.com | /HarveyCastroMD | #DrGPT

Disclosure

The author founded DR GPT™. The patient example in this article is hypothetical and does not describe a real person.

Related Posts