A few weeks ago, a study came out suggesting that eggs might be unhealthy — again. Many people rolled their eyes at what seemed to be a perfect example of how nutrition science can swing back and forth: “I don’t think I can read any more articles like this,” one reader wrote in a top-rated comment on the New York Times’ coverage of the study. “There have been contradictory reports on eggs, coffee, wine, etc. for so long now that there just doesn’t seem to be a true consensus on anything.” In response to JAMA’s Tweet about the study, one reader responded with “This again????” NPR asked: “Why do studies so often flip-flop from one answer to another?” and pointed toward a movement urging science to “embrace uncertainty.”
In the meantime, a different kind of answer has begun to emerge. Or at least a different question: What if it isn’t whether a given food is inherently good or bad for everyone, but rather how that food affects a given person, on an individual level?
In a recent op-ed for the Times, cardiologist Eric Topol, director of the Scripps Research Translational Institute, wrote about his experience in a nutrition experiment. For two weeks, researchers measured his food intake, his blood sugar, his sleep, his exercise levels, and the various species making up his gut microbiome. At the end, they filtered all the data through an artificial intelligence algorithm to come up with something like a personalized report card for the different foods he ate, based on how they interacted (or might interact) with his particular body and lifestyle, as determined by their likelihood to prompt glucose-level spikes. Some unexpected foods got A’s, like cheesecake and bratwurst, while seemingly healthy foods like oatmeal, certain fruits and vegetables, and veggie burgers got C’s. (This made me want to participate in the experiment so badly!)
As it becomes clearer that different foods do different things to different people, it also becomes clearer that “there is no such thing as a universal diet,” as Topol put it in his piece. There is no single best way to eat. What works for one person might not work for another. Which also means that there isn’t necessarily an objective answer to questions like “are eggs healthy,” either.
But then, another question: If there is no ideal universal diet, how do we figure out our own ideal individual diets? What should I be eating that’s perfect for me? The field of AI-guided diet is in its infancy, although it’s developing rapidly, as Topol details in his new book, Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. By this time next year, I might know for certain if almonds and apples are, in fact, the ideal foods for my own digestive fingerprint. Or something like that.
When might artificial intelligence be able to tell me what food I should be eating, based on my own individual bodily functions?
It’s doable now. At the moment there isn’t any commercial entity that’s doing it properly, so there aren’t any offerings that readers could jump on right this minute, but I mean, I did it, in the experiment that I wrote about, and thousands of people have done the experiment. It could be launched commercially — and it probably will be, as the little pieces come together.
But there’s a caveat.
Which is what?
We could individualize a diet that would be designed to keep your metabolic response ideal, in every respect. But then there are several questions after that. One is: Would that diet be palatable? And another is: Would it be worth it? Is it going to keep you healthy and promote your health long-term?
So, having an ideal metabolic response sounds good but hasn’t actually been proven to affect long-term health, yet?
Right. There are some things that are intuitive, but they’re not proven. We are going to figure these things out — there’s too much momentum now to not get there — but we’re not going to know about this for quite a while.
Someday, longer term, we’ll be able to see whether or not such an AI-guided diet — if people adhere to it, like in randomized trials — prevents outcomes like the development of diabetes, or heart attack, or strokes, and if it contributes to better survival. But that takes years and large numbers of people and rigorous studies, and that’s going to be an overhang that’s going to keep us in the zone of uncertainty for some time.
In the meantime, we’re going to know about things like your glucose levels, your triglyceride levels, certain lab tests. But that’s different from knowing whether following this individualized diet will prevent heart disease or diabetes — it’s going to take a long time to prove that.
So it’s like the beginning of an exciting new era in personalized nutrition, but where nothing is certain yet.
Well, one thing is certain: The idea that we should recommend a single diet for all people is ludicrous. And promoting the idea of a universal healthy diet is what countries and organizations have been doing forever.
For years we never knew that our individual responses to the exact same foods, in the exact same amounts, at the exact same times, were going to be so remarkably variable from person to person. But we’re seeing that now. It’s unquestionable. It’s replicated. The individuality of our response to food is unquestionable. But what remains is now getting this individualized recommendation of what you should and shouldn’t eat.
I know that the experiment gave you some unusual results regarding the foods you were told to eat and avoid. If there’s no universal diet, you’d think that eating intuitively would make sense, which is why it was surprising that many of the foods they recommended to you were not intuitively tasty to you.
No, they weren’t intuitive at all. But they checked out, later. [The foods that the experiment said would lead to glucose spikes did in fact lead to glucose spikes, when I checked later on my own.] There is something to this.
But so, for instance, they rated cheesecake as an A for me, personally. And I would never eat cheesecake. I’m a cardiologist; I’m allergic! No, I’m joking, but I mean — I also wouldn’t eat bratwursts, which they also said were an A+.
On the other hand, I really do like certain foods that got poor ratings, for me. One of my favorite vegetables is squash, baked squash, and that got a C, and I said, “What?!” And I like yogurt a lot, and that got a C as well. Another example — veggie burger. I don’t eat red meat at all, I haven’t for almost 40 years. So if I were going to have a burger, it would be a veggie burger, but that got like a C- rating. As did a lentil veggie burger. I said, “Wow.” I thought that was a healthy thing. And it isn’t necessarily unhealthy; it just means that I’m more prone to a glucose spike after eating it.
But then the question is: What do you do with that information? I haven’t yet made any resolution to adhere to this recommendation list.
What’s the reasoning, if you don’t mind me asking, behind not eating red meat?
Oh, for me, well I have a family history of colon cancer. On both sides, my mother and my father’s side. And I am very much in touch with the fact that there’s a relationship between red meat and higher risk of colon cancer. And then added to that, as a cardiologist, I’ve been telling my patients to limit their red meat intake for as long as I’ve been practicing, which is 35 years.
I’ve been eating less red meat lately, and I’m always curious why people make that choice.
Well, I don’t think red meat is particularly healthy, and it’s kind of interesting because on this output they had bratwurst being an A+ for me! Unfathomable!
And a veal cutlet, A+!
I review this a lot more in the book, but I think it sets up a quandary. We’re definitely seeing progress here, but we’re still a ways away from the point where people can really use this information to change their behavior. Because the only company currently offering a way to test some of this — Day Two — is in my mind not acceptable, since it only offers a gut-microbiome test, without the other tests.
I think there will be other entries to this field of testing diet with AI over the next year or two, though. And we’ll get there. You can almost tell that we’re headed into really sound recommendations on an individual basis. But then, also — is this going to be popular? I don’t know.
That’s such a no-brainer to me, because I’m dying to do all these tests. I’m almost embarrassed of how much I want some machine to tell me what to eat, based on an algorithm and millions of pieces of data. It’s so tantalizing.
I think that for any individual who really wants food to be medicine for them — who wants the things that they eat to actually promote their health — we are going to get there for those people. It’s a shaky field right now, but we are going to figure out what foods work for them. There’s going to be science here.
To do it, though, requires filtering an inordinate amount of data. For instance, you have the gut microbiome, with billions of pieces of data, you’ve got your sleep data, your physical activity data — exercise — you’ve got everything you eat and drink, you’ve got your medications. You know, other lab tests. So in order to get this machine learning model done for any person, it requires collecting all this data, from thousands of people. And that’s why we’ve finally made progress — we didn’t even have a way to analyze that kind of data until recent times.
And that’s the artificial intelligence.
Yeah. Since we’re onto food, and you’re into food, AI is like — AI has an insatiable hunger for data.
And humans have early satiety with data. No human can handle all this data, from all these different fields, to figure out what the right diet for each person is.
I think one of the big breakthroughs here was the glucose sensor. Four or five years ago, I tried a glucose sensor, and I saw these glucose spikes when I was eating. At the time, everybody thought that if you’re not diabetic, which I’m not, that you should never have high glucose — that it should always stay below 100 no matter what. Well, when I saw these spikes to 140, some years ago, I said, “Oh, this is horrible,” but no — at that time, so few healthy people had tested their glucose after they ate, using a continuous sensor, that we didn’t know that it’s normal for at least a third of people, and maybe even half, to have significant glucose spikes after they’ve eaten certain foods.
So again, these fields are just opening up. We didn’t have ways to get at this, back then. We didn’t have AI. We didn’t have sensors, we didn’t know about the gut microbiome. And all of this stuff is coming together. And what is exciting here is that we’re beginning to understand the underpinnings of our individual responses.
I’m so curious to learn more.
This conversation has been lightly edited for length and clarity.