Precision nutrition is about better tailoring diets and dietary recommendations to different people because one size certainly does not fit all, as I’ve written previously for Forbes. So to determine the best diet for someone all you have to do is figure out what’s going on with that person’s genetics, physiology, microbiome, body type, eating behaviors, stress, social influences, food environment, health conditions and all kinds of other stuff that affect nutrition and health. And you have keep track of how all of these things may interact with each other and change over time. No problem, right?
Not really. It can be really complex keeping track of and sorting out all these different things that are occurring in different ways at different levels for different people over different times and circumstances. That’s a lot of “differents.” These days, though, anytime you’ve got something very complex to sort out, you’ve got a potential friend in AI—meaning artificial intelligence.
One big challenge is that science has not yet even figured out how all of these different factors may interact to impact how a person’s diet may affect his or her health. Sure, studies to date have generated insights on how each of these factors may act separately and for certain types of people. But combining these insights is a different matter and a lot of gaps do remain.
That’s because a single traditional real-world laboratory, clinical or epidemiological study alone cannot account for, measure and keep track of all of the different factors and outcomes for all kinds of people. No matter how hard you try to design the “perfect” study, you will undoubtedly fail to include all types of people and measure every relevant factor and outcome.
Plus, even if you were to design the “perfect” study, you’d have to wait a long, long time to get all of the results needed. It can take years, even decades for the effects of nutrition to manifest as different health conditions. Anyone who ate like a garbage disposal and counted ketchup as a vegetable throughout his or her 20s will tell you that.
So if you really want to figure out how to do precision nutrition, you’ve somehow got to combine data from many different studies and fill in the gaps. You also want to find ways to extend the results of a given study to people who didn’t participate in that study and circumstances that were not covered. All of this can be way too complex for any given human or even a team of humans unaided to do.
Enter AI and cue the Randy Newman song, “You’ve Got A Friend in Me.” Such computer-aided techniques can keep track of many different things, combine different datasets in different ways and figure out how they fit together. These techniques can also determine how results from a single nutrition study might apply to differing conditions and situations, thereby elevating the utility and value of that study. And various AI techniques can do it quickly, much quicker than humans. These are just some of the ways AI can help achieve precision nutrition.
To understand how AI can do these things, you’ve got to first know what AI is. These days AI has become such a sexy term that people may use it without even really knowing what the term means, such saying as stuff like, “Hey, can you AI this?” AI is an umbrella term that basically encompasses any computer-aided technique that can replicate something that a human brain would normally do beyond simply following step-by-step directions. So an AI approach might assess situations or make decisions on its own. There are already many different types of AI approaches, methods and tools and the list continues to grow each year.
One way of classifying AI techniques is on a continuum of how these techniques are designed and operate. At one end are purely data-driven AI approaches. These are “top-down” techniques that start with a body of data and try to figure out patterns, trends and associations from this data. It’s a bit like how a statistician may analyze a set of data. But the AI algorithm can do it much faster and do many different analyses across multiple data sets simultaneously.
Let’s look at a theoretical example. A data-driven AI approach may analyze various data sets, slice the data in different ways and find that people who eat a certain food item tend to live longer. Let’s call this food item a “Swift Bieber,” a completely fictional term named after nothing in particular. The AI algorithm then may associate Swift Biebers with greater longevity but not explain why this association actually exists. It cannot really distinguish whether Swift Bieber consumption has some actual beneficial nutritional affect versus some kind of coincidence occurring. Maybe those who tend to eat Swift Biebers may also concurrently tend to eat another food item not captured in the data set that is actually doing the trick. Or perhaps people who have less stress are more likely to have the time and money to eat Swift Biebers. Swift Biebers could actually be a red herring, meaning something misleading or distracting rather than something made out of fish.
At the other end of the spectrum are mechanistic or explainable AI approaches. These AI methods attempt to recreate from the bottom-up what is actually occurring by recreating the actual mechanisms behind a process or decision. They’re deemed explainable because you know the specific reasons why a result was generated.
This is analogous to what scientists do when they design experiments in a laboratory to test what may happen. The difference is that the AI algorithm or model isn’t restricted to a physical laboratory and can serve as a “virtual laboratory” representing a whole person, a set of people, a population or an entire geographic area. The model can then run experiments in the “safety” of a computer in ways that would be too complex, too costly, too time-consuming, too impractical or even too dangerous to do in real life. The mechanistic AI tool could then use the results of these experiments to then determine recommendations, just like a human runs thought experiments in his or her head before taking action.
So, for example, a mechanistic AI approach might be to represent the different reasons why a person may choose to eat a Swift Bieber. It could also represent the different nutrients in a Swift Bieber, how they are broken down in the body, how these nutrients then affect different organs and then how this ultimately affects longevity. Then this AI model could then look at what would happen over time if different people were to eat Swift Biebers and decide who would benefit from eating Swift Biebers and how.
These different AI techniques along the spectrum can work together and be integrated as well. A purely data-driven approach can suggest associations (e.g., take a closer look at Swift Biebers) that can guide the construction of more mechanistic AI approaches (e.g., let’s figure out what Swift Biebers are actually doing to the body). Similarly, a mechanistic AI approach can help define where data-driven approaches are needed. Say when you are trying to represent the mechanisms by which a Swift Bieber affects the microbiome but can’t figure these out since there are no traditional studies that clearly teased out associations, patterns and trends. Therefore, it may be useful for for data-driven AI approaches to sift through this microbiome data.
Of course, one shouldn’t automatically trust anything that AI tells you. Just like a poorly-designed clinical trial or observational study can lead to misleading results, so can a poorly-designed AI approach. That’s why you’ve got to know what’s under the hood of an AI approach and understand its relative strengths and weaknesses. At the same time, no AI approach—just like no real-world study—will be perfect. Don’t let the perfect be the enemy of good and let the imperfections of an AI approach prevent you from using it out of risk-aversion.
Integrating more AI and other computer-aided approaches to make more precise recommendations is not completely new and has been done in other fields. Fields like meteorology, finance and aerospace engineering have long used computer-aided techniques to bring together and analyze complex data from different sources and generate more accurate insights and predictions.
So, while AI probably won’t go against some of the already established nutrition insights such as the value of eating fruits and vegetables, the field of nutrition is ripe for a change. There are too many people out there claiming that such-and-such-super-duper-just-eat-this diet works for everyone. But not everyone is the same and has the same circumstances, which is precisely the problem. Achieving more precision nutrition is not easy. but you’ve got a potential friend in AI. But like any potential friend, you’ve got treat it right and know what it can and can’t do.
