To hear more audio stories from publications like The New York Times, download Audm for iPhone or Android.
After 20 years of living with Type 2 diabetes, Tom Idema had given up hope of controlling his condition. He had tried many diets that proved unsuccessful and even considered weight loss surgery. When his employer offered him a chance to try a new dietary app that uses artificial intelligence to control blood sugar, he took it.
Mr. Idema, 50, sent in a stool sample to get his microbiome sequenced and filled out an online questionnaire with his blood sugar, height, weight and medical conditions. That data was used to create a profile for him, to which he added continued blood sugar measurements for a couple of weeks. After that, the app, called DayTwo, rated different foods according to how good or bad they might be for Mr. Idema’s blood sugar, to aid him in making better food choices.
After nearly 500 days using the program, his diabetes is in remission and his blood sugar levels have dropped to the upper end of normal. And even though DayTwo says the app isn’t aimed at weight loss, he’s gone from 320 pounds to 229 pounds. “I’m wearing pant sizes I haven’t worn since high school,” said Mr. Idema, who is an administrator at Central Michigan University in Mount Pleasant, Mich.
DayTwo is just one of a host of apps claiming to offer A.I. eating solutions. Instead of a traditional diet, which often has a set list of “good” and “bad” foods, these programs are more like personal assistants that help someone quickly make healthy food choices. They are based on research showing that bodies each react differently to the same foods, and the healthiest choices are likely to be unique to each individual.
Whether these A.I. nutritionists are ready for widespread use is still unclear, and there is very little research available from sources outside the companies selling apps. Users should be wary of overly broad claims that go beyond predicting how foods affect blood sugar.
But proponents say blood sugar is just the beginning and that artificial intelligence programs could target other aspects of metabolic health, such as obesity and heart disease, eventually helping to guide a person’s everyday meal choices.
How to make (artificially) intelligent food choices.
The DayTwo app uses an algorithm based on research by Eran Elinav and Eran Segal of the Weizmann Institute of Science in Israel, who co-founded the company in 2015. Last year, the company found that when they used their algorithm to match a diet to an individual’s microbiome and metabolism, it was better at controlling blood sugar than the Mediterranean diet, considered one of the healthiest in the world.
“Rather than measuring foods by caloric content and trying to come up with a ‘healthy diet,’” said Dr. Elinav, “you need to start measuring the individual.”
This technology is relatively new and only relates to blood sugar. The Mediterranean diet, meanwhile, has decades of research behind it and will likely remain the gold standard for healthy eating for years to come. Still, for people like Mr. Idema, A.I. like DayTwo’s can make it easier to maintain healthy eating patterns.
The app’s machine-learning algorithm can identify patterns and learn from data with human help. It analyzes data from different individuals’ blood sugar responses to tens of thousands of different meals to identify personal characteristics — age, gender, weight, microbiome profile and various metabolic measurements — that explain why one person’s glucose spikes with certain foods when another person’s doesn’t. The algorithm uses these observations to predict how a particular food will affect one’s blood sugar and assign each meal a score.
The system can’t yet take into account the candy bar someone had two hours ago — but users can play around with food combinations to change the score for each meal. For example, the app gave macaroni and cheese — one of Mr. Idema’s favorites — a low score, but he was able to improve it by adding protein. That’s because adding protein or healthy fats can temper the blood sugar spike from a carbohydrate-heavy meal like macaroni.
“I thought they were going to say, ‘Oh my gosh, you’ve just got to become a salad eater, and that’s not been the case,’” said Mr. Idema.
DayTwo, which is currently only available to employers or health plans, not consumers, is one of a handful of A.I.-based apps recommending healthier meal options. Another company, ZOE, also generates meal scores and is available directly to consumers for $59 per month. ZOE’s algorithm uses additional data, such as blood fat levels, in addition to microbiome and blood sugar tests. The algorithm was able to predict how a person’s blood sugar and fats respond to different foods in a large 2020 study led by one of the company’s founders, Dr. Tim Spector, a professor of genetic epidemiology at King’s College in London.
Currently these algorithms mostly focus on blood sugar, but newer versions will incorporate more personal data, and, in theory, recommend diets that reduce cholesterol, blood pressure, resting heart rate or any other measurable clinical indicator.
“Bringing in all these different data types is very, very powerful, and that’s where machine learning kicks in,” said Dr. Michael Snyder, a genetics professor at Stanford University who helped found the health start-up, January.
The field of personalized nutrition is still in its Wild West phase, and experts say it’s important to sort through the hype. Many companies are willing to test your microbiome and offer A.I.-driven dietary recommendations — as well as sell you supplements — but few are based on scientifically rigorous trials. Last year, uBiome, which made one, was even charged with fraud. In general, the more broad-ranging the health and weight loss claims the companies make, the less reliable the evidence to support them.
“I think it is all overhyped right now, unfortunately,” said Dr. Eric Topol, a cardiologist and the founder and director of the Scripps Research Translational Institute.
The data used by apps like DayTwo and ZOE also only capture a fraction of the interplay between the gut microbiome, our metabolism and diet. There are certainly a lot more factors, like genetics, that affect metabolism and are ignored by current A.I. programs.
“It does not tell you the whole story, and just optimizing around glucose is not going to be enough to create the perfect diet for you,” said Dr. Casey Means, co-founder and chief medical officer at a digital health company called Levels. A.I. apps could nudge users into eating foods that are good for preventing blood sugar spikes and diabetes, but may be unhealthy in other ways.
For instance, when Dr. Topol tried out the DayTwo app, its recommendations for controlling his blood sugar — such as eating spinach and raspberries — were high in oxalic acid, which could have induced kidney stones. That’s because the app didn’t take into account his pre-existing risk for the condition.
Additionally, restrictive diets are increasingly seen as a bad way to change eating habits and often backfire. But many experts hope personalized A.I. apps will be easier to follow and build better long term behaviors.
For now, these apps could assist nutritionists with meal suggestions, but aren’t going to replace them, and both ZOE and DayTwo have regular virtual checkups with a dietitian or nutritionist built into their programs.
According to Dr. Topol, larger and longer-term studies that incorporate more layers of data, such as sleep, exercise or stress, into the algorithms could make these programs more precise and accurate for each individual. They could also help people see how short-term responses, such as post-meal glucose spikes, influence long-term health.
What we don’t know is how or if the day-to-day improvements translate to long-term health. Dr. Topol said of A.I. diet programs. “Can you prevent diabetes? Can you prevent heart disease and other chronic diseases?”
These larger studies are coming. The National Institutes of Health’s Nutrition for Precision Health research program began a multiyear study in January to develop algorithms to predict individual responses to foods.
But for Mr. Idema, the effects of personalized diets are already tangible, most recently when his improved blood sugar levels allowed him to enjoy his daughter’s birthday cake. “I had the glucose monitor out at the time, and I stayed well within range, so my body handled it just fine,” he said. “So I’m in such a much, much better place now, and in my mind this program definitely saved my life.”
Sandeep Ravindran is a freelance science journalist based in Bethesda, Md.
Audio produced by Kate Winslett.