Every morning for about 100 days, one reviewer's first task was to feed something that does not eat. The thing in question is a digital twin, and its appetite is for data: body composition from a smart scale, a log of the lemon water used to take morning medications, and on Mondays a blood pressure reading after some calming breaths. Only then does the human get to think about breakfast.
The arrangement is odd, but the results are more interesting than the oddity. Labs improved. Weight did not. And the person doing the feeding came away with sharper questions about what happens when this technology leaves the low-stakes world of food and enters the high-stakes world of medicine.
What a digital twin actually is (and is not)
Despite the name, nobody gets a robotic clone. In this case the twin is an AI-powered, color-coded food database that predicts how your glucose will respond to what you eat. Red means not so good. Orange means so-so. Green means ideal. It is built from the data you log, plus readings from a continuous glucose monitor (CGM).
In medicine and sports, the broader idea is newer and more ambitious. A digital twin is a virtual model of a person assembled from labs, wearable data, images, genetic profiles and similar inputs, then updated continuously. Clinicians could then use AI to see how a patient might react to a specific drug or treatment by trying it on the twin first. In theory, that could also help providers assess injury risk, predict how a disease will progress, and plan preventive care while sparing the real patient adverse effects.
That is the dream of personalized medicine in one sentence. Metabolic coaching, the reviewer argues, is perhaps the lowest-stakes, highest-reward version of it that can be tested in the real world today.
The setup: a very large box
The twin was "born" in June, when a big box from Twin Health arrived. Twin Health is a program meant to help people "heal" their metabolisms, which covers things like weight loss, diabetes management and better sleep. You cannot simply download it. The service is currently available only through corporate or private health plans, and it is typically aimed at people with chronic metabolic issues.
The box held the raw materials for building a twin:
- CGMs: 90 days' worth of Dexcom G7 sensors.
- Home devices: a smart scale and a smart blood pressure cuff.
- Supplements: three bottles of multivitamins and magnesium glycinate.
- Existing gear: the user's own Apple Watch.
A trip to Quest Diagnostics followed, where a phlebotomist took blood for further analysis.
A few days later came a Zoom call with a Twin Health coach named Amanda (a human, as the reviewer is careful to note) and a doctor. The numbers were not a surprise to someone who had spent a year testing CGMs as a non-diabetic. The liver was inflamed, which fit with a prior diagnosis of moderate nonalcoholic fatty liver. Insulin resistance was clearly present, even though HbA1c and fasting glucose looked ideal. The pancreas appeared to be working overtime, though the count of insulin-producing beta cells looked fine, at least for now. Triglycerides were high, and the remaining lipids were normal.
Asked to name a goal, the reviewer said it plainly: take fewer medications to keep a middle-aged metabolism running. Amanda logged it, booked a follow-up in two weeks, and asked for as many logged meals as possible during the first 30 days. A second blood draw was scheduled for the 90-day mark.
A morning with the twin
The daily routine is simple enough to describe and tedious enough to resent. Log the body data. Log the water and the medications. Consult the twin about breakfast.
Take a bowl of protein Cheerios with milk. Eaten alone, the twin rated it orange. Followed by the daily walk, it became green. So that is what happened. The meal gets logged, and a few hours later the app generates an insight based on how glucose actually behaved, which makes the twin a little more accurate.
There is some novelty in checking the color of random foods. A cup of pineapple came back orange. A turkey hot dog with ketchup and relish came back green. A bowl of soondubu jjigae with white rice landed on green or orange depending on whether the reviewer was active afterward, which the reviewer took as personal vindication against low-carb orthodoxy. (The phrasing was considerably more dramatic and involved rice and cold, dead hands.)
That last example hints at a pattern. Walking after a meal was often the difference between one color and the next. Today's test case, a protein waffle with a tablespoon of peanut butter and honey, plus pineapple, came back orange. A walk, once again, can turn it green.
The unmasking: the twin is hungry, and so is the bill
The pitch for a digital twin is that it does the hard thinking for you. The practice is that you do a great deal of work so that it can. Over the months, the reviewer noticed how hungry the thing was, and how draining it felt to keep feeding it. This came from someone already accustomed to tracking and reviewing piles of health data. In the check-ins with Amanda, she mentioned that tracking fatigue is frequent feedback from Twin Health users.
There was also a quieter discomfort. The reviewer found echoes of Kazuo Ishiguro's novel Never Let Me Go, in which clones are raised as organ donors. The comparison is not one-to-one, and the reviewer admits as much. An AI-powered twin cannot eat, die or feel pain. The unease is probably anthropomorphism, a feeling that testing food on a virtual copy of your metabolism resembles a jester-like food taster in a medieval court. Still, it was the kind of thought one has when a spreadsheet starts to feel like a dependent.
Nor were the predictions revelatory. The reviewer had tested too many CGMs and AI nutrition apps to be surprised by what the twin said, though they allowed that the insights would likely be useful to someone new to metabolic issues.
What actually changed
Behavior shifted, in expected and unexpected ways. The expected one: having to log food meant avoiding unhealthy treats more often than not, mostly because nobody wants to be scolded by an app. The unexpected one: the app was not the main motivator.
The twice-monthly check-ins did the heavy lifting. Stating progress aloud to another human who is not family kicked the reviewer's inner overachiever into gear. After months of feeling too weak from medication side effects, the reviewer recently ran an uninterrupted 5K and returned to strength training, helped along by some cajoling from Amanda.
The numbers after 100 days:
| Measure | Result |
|---|---|
| Metabolic health score | Up nine points since joining, per Amanda |
| Liver enzymes | All but one back to normal, a year after a doctor's eyes bugged out at them |
| Triglycerides | Down 70 mg/dL |
| hs-CRP (inflammation marker) | Significant drop |
| Insulin resistance | Still present, but less so |
| Weight | No meaningful loss |
| Medication tapering | Did not go as smoothly as hoped |
A fair reading comes with a caveat the reviewer supplies themselves. The liver numbers had been improving slowly before the twin arrived. They fell faster during the trial, but the trend was not created by it. So the twin gets partial credit, and the human coach, the walking and the renewed training are all hard to separate from it.
Why low stakes is the whole point
"I can't really argue with results like that, can I?" the reviewer asks, and then answers: actually, yes.
A digital twin runs on trust. In metabolic use cases, that trust is cheap to extend. If the twin gets your weight loss, diabetes management or sleep wrong, failure is not immediately life-threatening, and other medications and avenues exist.
Personalized medicine is a different beast. Researchers are exploring twins well beyond metabolism, including in cancer care and research, where the potential is large. Chemotherapy is grueling, and a way to test options virtually first would be a real prize. But the challenges are large too, starting with the burden of trust around data privacy. A twin's accuracy depends on a continuous stream of high-quality, sensitive information about you, which then informs future decisions.
The reviewer's questions are the right ones, and none of them has an answer here:
- If three months of logging felt overwhelming in a low-stakes setting, how does that translate to a high-stakes one?
- How does a digital twin exist once treatment is over?
- How does that data "live on" if treatment is unsuccessful?
Round two
The twin will get another three months. The reviewer has not given up on tapering off some medications and could use all the help available in bringing liver labs fully back to normal. There is one concession: Amanda says there is no need to be as militant with the data now that a solid foundation exists. Lower stakes, again, mean more flexibility.
Which may be the most honest feature of the whole experiment. The twin learned the reviewer's metabolism, and the reviewer learned that a model this hungry can afford to be fed a little less. For breakfast, in the meantime, there is a protein waffle, and a walk to follow.