The message arrived while I was packing for a trip. A personal AI system built around my health information had noticed a change in my recovery pattern and suggested that I reconsider traveling. Nobody had asked it anything. I was preparing to leave and doing what I suspect many busy physicians will recognize: treating the schedule as fixed and assuming my body would cooperate.
So, I changed my plans. That decision deserves an explanation, but not the dramatic one a technology headline might suggest. The system didn’t diagnose an illness. I can’t say that staying home prevented a medical emergency, and there’s no counterfactual version of that trip for me to examine.
What it did was more modest. It made me pause long enough to question the familiar argument I had with myself: that I could push through.
As an emergency physician, I’m comfortable moving fast when a patient needs immediate attention. Paying closer attention to my own health has called for a different discipline: noticing a pattern before I’ve decided it’s inconvenient.
That necessary shift is why I built what I call a personal health twin, because the most useful lesson has less to do with what artificial intelligence knows than with what I’m willing to notice.
What I Mean by a Health Twin
The term needs a boundary. My system isn’t a complete digital replica of my body, and I don’t present it as a clinically validated tool for predicting disease or simulating treatment. Recognizing these limits helps prevent over-reliance and ensures responsible use.
I use it to organize selected health information, including sleep, activity, recovery, and wearable trends, while being mindful of privacy and data security. Its job is to draw my attention to changes and help me weigh them in context. It doesn’t diagnose, prescribe treatment, or replace a physician.
This distinction matters beyond my own project. A scoping review published September 30, 2025, examined 149 studies describing human digital twins and found that only 18 met the researchers’ application of the National Academies’ criteria, which emphasize personalization, dynamic updating, and predictive capabilities.1 The vocabulary is a moving target, and it’s moving faster than any agreement about what these systems actually do.
For a physician sizing up any product with “twin” in its name, I’d start with ordinary questions. What information goes in? What happens when some of it is missing? What does the output claim to mean, and how was that claim tested?
An impressive name can’t answer those questions. Neither can an attractive dashboard.
The Difference Between a Measurement and a Decision
In my presentations, I describe an isolated laboratory value as a photograph and a series of measurements as a film. The comparison holds up, but a film still needs someone to watch it critically.
Consider a hypothetical patient whose wearable reports worsened recovery after several disrupted nights. Before drawing any medical conclusion, I’d want to know about symptoms, travel, training, medications, and how consistently the device was worn. I’d also want to separate the device’s raw measurements from its proprietary interpretation of them.
My standard for a useful summary is simple. Show what changed, over what period, and how much information is missing. Keep the observations apart from the suggested explanations and don’t let an uncertain pattern harden into a confident diagnosis.
Physicians are already running into these questions. In a survey of 2,222 physicians across six countries, 97 percent reported reviewing wearable data in some capacity, yet no country reported integration rates above 6 percent. Workflow, reimbursement, trust, and liability remained important barriers.2
To me, that gap argues for making information easier to evaluate, not for asking clinicians to absorb an endless stream of it. A patient shouldn’t need to arrive with 40 screenshots, and a physician shouldn’t have to reconstruct the timeline during a short appointment.
A better conversation is the real destination. The dashboard only gets you to the door.
What Actually Changed for Me
The travel decision is the clearest example from my own experience: instead of waving off a change in recovery, I reviewed it, rested, and adjusted my schedule accordingly.
That’s a behavioral observation, not proof of a health outcome. I’m not presenting a controlled experiment, and I’m not claiming that this system lowered my cardiovascular risk, reversed my biological age, or added years to my life.
Oddly enough, I think that restraint makes the story more useful to physicians. I can describe my own experience accurately without dressing it up as evidence for a product.
What interested me was the gap between knowing that recovery matters and actually making room for it. The system didn’t supply a new medical principle. It pulled an existing concern back into view at the exact moment I was ready to shelve it.
It also gave me a practical test for future alerts. Did this message help me understand something, or did it just make me anxious? Could I name a sensible next step? Would I reach the same decision after reviewing the underlying numbers without the AI’s wording in front of me?
Those questions now shape how I judge usefulness. Any emergency physician who has lived through alert fatigue knows that more alerts don’t automatically mean more benefit, and a persuasive explanation isn’t necessarily a more accurate one.
A Framework for Using the Information
I call my approach SCAN: See, Connect, Act, and Navigate. It’s a way to organize questions, not a validated clinical protocol or a substitute for individualized care.
See the signals. Start with a specific question rather than a shopping list of measurements. What are you trying to understand, and what information would actually help? For me, the starting point is a pattern over time, not a contest to collect the most data.
I wouldn’t turn this into a recommendation that everyone order broad laboratory panels or buy several devices. Testing should answer a clinical question, and a good system should admit when the available information cannot answer it.
Connect the context. Ask what else was happening during the period being summarized. Was the schedule different? Were measurements missing? Did symptoms change? A useful summary makes it easy to inspect the observations before accepting anybody’s explanation of them.
Keep the line between observation and inference visible. “These recorded values changed” is a very different statement from “This is why you feel unwell.”
Act conservatively. In my travel example, the action was rest and a schedule change. I didn’t treat the alert as a reason to order an extensive workup or start another supplement.
Certain situations call for responses that fit the person and the concern. I wouldn’t use a consumer AI alert alone as the basis for a medication change, a diagnosis, or a decision to ignore symptoms.
Navigate with a clinician. Bring organized information and a clear question to someone who can assess the full clinical picture. The system should support that relationship, not imply that a personalized summary has made it unnecessary.
Emergencies sit outside this framework entirely. New or severe symptoms need prompt clinical assessment, and a reassuring score should never become permission to postpone urgent care.
The Physician Still Has to Think
Personal experience can make me enthusiastic about a tool. It can also make me less skeptical of its next answer, and that’s the part I watch.
In a randomized study published April 23, 2026, 44 physicians who had completed AI-literacy training worked through clinical vignettes with access to either error-free AI recommendations or recommendations containing deliberate errors. Those exposed to the erroneous recommendations performed worse on diagnostic reasoning.3 The study used simulated cases rather than patient care, and it doesn’t tell me how often my personal system would be wrong.
What I take from it is simple: Experience isn’t immunity. Knowing medicine, and knowing that AI makes mistakes, doesn’t remove the need to examine each answer on its merits.
My own habit is to state the observation before I read the explanation. Then I ask what would make the proposed interpretation less likely, and whether the next step would still make sense if the explanation turned out to be wrong.
The same discipline applies when a patient brings me an AI-generated health summary. Rather than endorsing or dismissing the document wholesale, I start with the patient’s concern, check the relevant underlying information, and separate useful observations from unsupported conclusions.
That approach leaves room for an engaged patient without granting the software authority it hasn’t earned. It also keeps any disagreement focused on the claim rather than on the person who brought it.
Privacy Belongs in the First Conversation
A personal health system becomes more revealing the more information it pulls together. Before recommending one, I’d want clear answers about where the data is stored, who can access it, whether it’s used to train models, and how it can be exported or deleted.
Those are questions to investigate, not protections to assume. I’d also ask what happens when an account is compromised, a company changes ownership, or the user simply stops paying.
Publishing a personal story works by the same rule. Sharing what I learned doesn’t require publishing my complete records, account details, or family information, and a good example discloses only what the reader needs.
I’d rather explain a limited system accurately than describe an ambitious one whose data practices I couldn’t defend. Trust should rest on what can be checked.
A Better Question Between Visits
My interest in this work is ultimately about healthspan. In other words, I want to stay capable of doing the things that matter to me, and a score isn’t that goal any more than a more elaborate personal dashboard is.
The question I ask audiences in my METAL presentation is more personal: Who do you want to be capable of showing up for in 10 years?
This question reminds me why paying attention matters in the first place, and no personal health twin will answer that for me. What I want from the system is narrower: Organize the information I can use, flag uncertainty openly, and leave room for the parts of a life that a sensor will never capture.
My experience hasn’t convinced me that medicine should become a continuous stream of automated instructions. It has made me more interested in a quieter possibility: Using well-organized information to help a person pause, ask a better question, and make a considered decision.
When that message arrived before my trip, the decision was still mine. What changed was that I stopped long enough to make it deliberately.
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.
To learn more visit HarveyCastroMD.com | /HarveyCastroMD | #DrGPT
