Isabelle Hoffmann, the CEO of a food scanning company called Tellspec, pours two delicate cups of organic keemun tea and sets a jar of raw honey in front of me on her kitchen counter. It reminds her of a conversation she had with a food analysis lab scientist. “He said that in 42 years of working there, there is not one time that he has tested pasteurized honey and not found acrylamide,” she says in a deep Brazilian accent that makes her sound authoritative in the way some British accents always sound polite. “It is a side effect of the pasteurization.”
Her heels clunk against the wooden floor of her apartment as she walks to her desk and returns with her laptop and a pair of blue-lined reading glasses. “Okay, here we go,” she says, opening a database that she calls the “TellSpecopedia” and selecting “acrylamide” from an alphabetized list of food components.
Someday, if her business goes as planned, we won’t need the laptop. We will use a handheld device, maybe even our phones, to scan our food, and an app will notify us about the presence of chemicals like acrylamide.
“Carcinogen,” she begins, counting off on her fingers the possible side effects of ingesting the chemical. “Reproductive organ disruptor, behavior effects, fetal development disruptor.”
“That’s the research, there,” she adds, pointing to a list of hyperlinked medical journal citations at the bottom of the page.
We both agree it would be better to not investigate too thoroughly the chocolate croissants Hoffmann has wrapped in tinfoil to warm in her oven. (“You can’t be too orthodox, right?” she says, popping a few blueberries into her mouth with one hand and adding, “they probably have trans fats.”)
There’s a lot we don’t know about food and our health. While there are traditional medicinal food practices like Ayurveda and that guy who is hawking turmeric everything at your farmers’ market, scientific research–and funding–tends to focus mostly on how to reverse diseases, not how to prevent them. “We know diet is a major factor for disease risk, but we really need to hone down on what individual components of diet are related to disease risk,” says Chirag Patel, a Harvard researcher who uses big-data techniques to learn about how environmental exposures, like what we eat, relate to disease and health.
Hoffmann’s company, Tellspec, could one day help refine that connection between diet and health.
Startups like 23andMe have, in hopes of learning more about the relationship between genes and diseases, started collecting a giant database of its customers’ DNA tests. Other companies, like uBiome, have begun similar projects around sequencing customers’ microbiomes (all the gut bacteria that helps us digest stuff). Tellspec wants to provide the same type of data for food, so that we can better link what we eat with how we feel.
And unlike our DNA or the bacteria in our gut, what we eat is something we can easily control. Imagine going to your doctor and hearing something like this: “We know what your microbiome looks like, we know what your genome looks like, and we know that people who have similar traits to you and ate this food often got this disease. We also know these people were more likely to avoid disease when they ate this other food. So, here’s how I recommend you change your diet.”
That’s Hoffmann’s dream.
The problem is that building a food-scanning device that can both correctly identify your food and use that data effectively is about as difficult as it sounds.
Hoffmann, who drinks a glass of goat milk every morning and gives employees affectionate nicknames like “whiz kid,” has barreled full-force across various industries and roles without hesitation. “As long as you believe you can solve it, as long as you believe you are intelligent enough to solve it, I think you can,” she says. “It may be difficult. And some people don’t have this attitude.”
Hoffmann’s background suggests that she believes in the power of positive thinking. She grew up in Portugal, but moved to Toronto as a college freshman to study astrophysics. She says her English wasn’t good enough to complete that degree, so she switched to math. Eventually she completed a PhD in mathematics education at the University of Toronto (though she rarely mentions it, saying she prefers to present herself as a businessperson and entrepreneur).
In the ’90s, after realizing her son’s playtime stories would make good interactive scripts, she started a computer games company called Hoffmann + Associates (H+A for short) that was soon one of the fastest-growing businesses in Canada. When that company–which had also acquired a business that produced games and medical reference CD-Roms–burst in the dotcom bubble, she simply changed tracks. “I thought, no more technology,” she says. “I want to do medicine. I started taking all sorts of courses. And in the course of this, I started meeting all the doctors I was working with in these CD-Roms.”
One of them asked her to help start an anti-aging clinic in Beverly Hills, which she did. That led to a genetics business, called GenoSolutions, which she launched in Portugal. Basically an earlier, more expensive, version of 23andMe, it had trouble finding customers. And that’s how Hoffmann came to accept a non-paying job as South European president of the World Academy for Anti-Aging and Preventive Medicine. The job entailed producing preventive medicine conferences, which she thought would be good promotion for her business.
Tellspec sprouted from all of these experiences–the medical conferences that gave her an encyclopedic knowledge of studies showing relationships between food and health, the games company that gave her experience with technology, and the belief that she could do anything she imagined–but its impetus was a more personal project.
After Hoffmann moved, in 2011, into a Victorian house in Toronto, her teenage daughter developed terrible hives and swollen lips. Her blood pressure was so low that she needed to use a rolling office chair to get to the bathroom at night without falling, and she couldn’t go to school. “I was told, every time I went to the hospital, go back home, she has a viral infection, and she’ll get better,” Hoffmann says. “But she didn’t get better, she got worse.”
Doctors suggested Lyme disease, but all of the tests came back negative. As her daughter took a buffet of pills to treat her symptoms, Hoffmann wrote to all of the medical doctors she knew. Finally someone recommended a clinic in California that combines conventional and alternative medicine, where tests revealed that her daughter was allergic to Penicillium Aspergillus. “I said, but she’s not taking penicillin,” Hoffmann says. “She’s not taking antibiotics.”
When she had her daughter’s bedroom tested for mold, 65% of the spores were Penicillium Aspergillus.
“I thought, maybe I have taken this environmental medicine thing too far, and I’m insane,” Hoffmann tells me while we’re walking between her apartment and the Tellspec office. “But I moved her into that building,” she points to a new condo building nearby, “and within three months, she was walking again.”
Shortly later, Hoffmann founded the second company inspired by one of her children.
“The idea of beaming your health up, the idea of getting rid of everything that is not good for you, and making your health super optimized,” she says, “That was the idea.”
Tellspec’s offices are located on a quiet residential street in Toronto, in the peak-roofed top floor of a townhouse. A hardware engineer and software architect share one nearly bare office. Two machine learning specialists, who are developing algorithms that can translate the device’s scans into recognition of food components, tap away in another another almost identically vacuous space. But the third office looks like a pantry.
There is bread everywhere: bread in freezers. Bread in the fridge. Dozens of homemade cakes zippered into plastic bags and lined up like soldiers on the freezer rack. Boxes of English muffins and crackers and pastas stacked on tables and shelves.
This is ground zero for Tellspec’s database of food.
The core of Tellspec’s technology, spectroscopy, is used for everything from analyzing blood to hunting for water on the moon. Techniques vary, but here’s the concept at its most simple level: When you shoot light at something (or, in the moon’s case, the sun does), the light that reflects back will be different depending on the shape of the molecules it just hit. A spectrometer can detect those differences by counting and sorting photons by wavelength, and a computer can decide whether or not a certain substance is present by matching that “light fingerprint” against a library of known light fingerprints. So let’s say you scan a piece of bread with a spectrometer. The result won’t say “bread.” It will say, “this substance has gluten and this type of protein and these types of sugars.” But it will only be able to say this if you’ve provided a sample of what reflected light from gluten, these types of protein, and these types of sugars look like. You need to provide data to match against.
Large spectroscopy databases exist for explosives and pharmaceuticals. But there is no existing spectroscopy database for food, which is why Tellspec’s offices are currently packed with bread.
Among all this gluten (and, according to the packages, “gluten-free”) food, two “food testers” (that’s the official title) wearing white lab coats sit at laptops on rolling chairs. One of them places a cracker over a square window of light on top of a shoebox-sized contraption. The Tellspec prototype is inside of the box, and data it collects from scanning the cracker light pops up on a laptop. Eventually those scans will be used by Tellspec’s machine-learning algorithm, along with a chemical lab analysis of the same type of cracker (which Tellspec uses instead of labels for defining what it is scanning into the database), to recognize the cracker’s ingredients when it sees them again.
Despite the many slices of bread in the room, this process is no picnic.
There are hundreds of different types of bread, all with different ingredients and seeds and textures. The outside of the bread might be a different consistency than the inside of the bread. Tellspec isn’t necessarily interested in telling people they’re eating bread. But it does want to tell them what core nutrients they are eating. So it has to scan those nutrients in all states to tell its machine, “gluten can look like this” and “gluten can also look like this.”
Then there’s toast. Because cooking molecules changes their composition, Tellspec will need to scan all of those bread samples as toast in order to teach its machine to recognize the ingredients accurately.
And that’s not the end of it. There is also buttered toast. And toast with jam on it. And the fruit-heavy spots of jam compared to the clearer spots of jam.
“And we’re only talking about bread,” said one person familiar with the science of spectrometry about Tellspec’s plan to build a food database. “What they’re talking about is quite a big feat.”
Wow, I say, when I realize the limitless scans the people in this room have ahead of them. You could be here until the end of time.
“Yes,” Neeshma Dave, who has a PhD in bio-analytical chemistry and is choreographing this whole operation, admits without flinching. “I suppose you’re right, it could take quite a while.”
There are two pieces of art in the Tellspec office. One is a gift from a woman who works with the Chinese government, which Hoffmann says is interested in food safety applications for Tellspec. It’s a roll of parchment that says, in broad-stroked calligraphy, “With hard work, you can accomplish your dreams.” Another, a gift from Hoffmann’s boyfriend, is a tiny plaque painted with a famous quote from Walt Disney: “It’s kind of fun to do the impossible,” it says.
Hoffmann is, of course, quick to point out why the quote is appropriate for her office. “How many frames do you have to draw to make these things animated?” she says about Disney’s cartoons. “It was too much work, wasn’t it? People have to draw so many frames in a second. Imagine. And today this is done with computers, no different from creating databases.”
When Tellspec first launched an Indiegogo campaign, Hoffmann didn’t quite know what she was getting into. She had a partner then, a professor of mathematics at York University named Stephen Watson, who has since left the company (Watson declined to talk to me for this article, saying his lawyer advised him against it). He had started to develop the learning algorithms for identifying foods using spectroscopy while he was with the company, but Tellspec hadn’t begun building the device itself.
The founders picked the wrong type of spectroscopy. They were overly optimistic about both their deadline and their sleek computer-mouse-sized concept design. They said beta testers’ scans would build the initial food database, without understanding the precision required to guarantee accurate results (the reason that, unlike Tellspec’s competitor Scio, the company ultimately decided to focus solely on food). And their demonstration video, filmed with a 3-D-printed dummy prototype, seemed to falsely represent a finished device. A few Indiegogo watchdogs picked up on these mistakes early. “If you feel like sending money to Canada hoping for the next great thing that promises to revolutionize the food industry, do it at your own risk,” one wrote, “But don’t be disappointed if you find out you sent money to a science fiction campaign.”
After raising $386,392 on Indiegogo and a seed round of funding, some of Tellspec’s ambitions still seem fanciful. The spectroscopy machines used by businesses and scientists are clunky large pieces of lab equipment that typically cost thousands of dollars and can cost as much as $20,000. Tellspec wants to make a handheld version priced for consumers, initially at about $500, that is wireless and portable. But cheaper and smaller spectrometers gather much less information than their beefier counterparts, and with the current technology, there is no way, within Tellspec’s size and price requirements, to identify food ingredients that initially excited many of its backers. Tellspec’s first version can’t detect pesticides or trace amounts of nuts, for instance.
Since launching its Indiegogo campaign, however, the startup has hired employees and contract workers with PhD.s in bio-analytical chemistry, applied mathematics, computer science, and educational psychology among them. It has created a partnership with Texas Instruments, which will supply a chip for the device, added 250,000 scans to its food database, and produced alpha and beta versions of the scanner. Its working prototype can identify calories, macronutrients (fats, protein, and carbohydrates), and select allergens like gluten.
Cast in another light, the same ambitions that make Tellspec seem naïve can instead look optimistic. Brave, even.
“I think it is possible,” says Rishikesh Pandey, a postdoctoral associate at MIT’s Laser Biomedical Research Center who works in spectroscopy. “In the last decade, we have made steps and progress in terms of having a more detection capability, having a smaller detection chip, having more sensitivity that can detect a signal.”
“It’s possibly possible,” says Sam Panariello, an OEM sales specialist who works with a company that designs and manufactures spectrometers called B&W Tek. “Maybe it can’t be done with current technologies, but new forms of spectroscopy are born pretty often. You could say with what we’re doing now it’s not possible, otherwise we would be doing it.”
“None of this is difficult,” says David Creasey, the marketing vice president of Ocean Optics, the spectroscopy company helping Tellspec build the device. “It just requires time, money, resources, and someone strategically wanting to do it.”
Meanwhile, Tellspec’s ultimate vision of using data to help prevent disease lines up with research exploring how our “exposome”–everything that we experience in our lifetime, from food to exercise to lifestyle–connects with our health.
There are only a certain number of hypothesis that scientists can test, and they will never be able to study how every food affects every type of person. But if some organization were to aggregate data from a food-tracking source like Tellspec with other data sources like 23andMe, uBiome, and even activity trackers like the Fitbit, they could start to make something that looks like a GPS map of health data. “It doesn’t have to be entirely causal,” says Dr. Joel Dudley, an assistant professor at Mount Sinai School of Medicine who is studying the relationship between microbiome, genome, and food allergies. “We don’t have to know why. It’s just, given everything I know about you, this is where you are in this GPS-coordinate universe of this health data, and I can see that when other people were in this location, when they reduced their carbohydrate intake, they went down this path and moved to this space, which seems to be a healthier space, and when they drank high-sugar soda, they went to this space, which seems to be a less healthy space with a higher probability of diseases.”
We are not, he cautions, nearing the ability to do this now. We need more data. And we need someone to take the developing silos of data–DNA, microbiome, activity, and food–and integrate them in one data universe. But it’s a track he sees as worth pursuing.
“I can tell you, if you’ve seen a late onset Alzheimer’s brain on the genetic level, you definitely want to prevent someone from getting into that state,” he says, “it’s a heck of a lot easier than trying to get them out of it once they’re there.”
Tellspec’s current beta unit is a black plastic contraption about the size of a guinea pig that has recently been screen-printed with the company’s logo.
Four Tupperware containers of bread sit next to the table beside it. They are labeled gluten-free one, gluten-free two, gluten one and gluten two. This is it: the demo I’ve traveled to Toronto to witness.
And indeed, when Tellspec’s VP of hardware engineering, Nenad Debeljacki, passes a slice from each box in front of it, the dashboard on Tellspec’s beta app changes to reflect calories per gram, whether the bread contains gluten, and whether it contains different fructose, glucose, or maltose. The label doesn’t have this information, Hoffmann points out. It just says “sugar.” Mark Bloore, Tellspec’s senior software architect, clicks on “glucose” in the Tellspec app, and I learn that it’s the body’s preferred energy source, and that its consumption leads to higher levels of blood glucose than consumption of sugars such as fructose.
So what does that mean for my health?
After watching the FDA intervene when 23andMe suggested people were predisposed to certain diseases based on their genes, Hoffmann has been careful to make sure Tellspec is not presented as a medical device. Instead, the app will ask people how they feel after each scanned meal. Were they bloated? Did they get hives? A headache? Over time, the app may make suggestions about what types of food a user is sensitive to, not unlike a restrictive diet.
Tellspec does not, at least initially, plan to collect the kind of information about its customers’ disease histories that 23andMe does, so it can’t necessarily be the company to connect the dots in the GPS map Dudley envisions. “We’re missing all the people who are going to put these data streams together,” he says. “There’s not a single stream of this data that is going to be totally predictive. We really need someone who is going to build an effort that will take the genome data, the microbiome data, the food scanner data, and other wearable data, things like that, and run studies that will build a model of how we put this data together.”
Researchers like Dudley and Patel are most immediately excited about the potential to use something like Tellspec’s gadget for collecting their own data. Current studies rely on questionnaires that people fill out about what they eat (e.g., did you eat foods with high fat?), an approach that is often, Patel says, “”fraught with problems like accuracy and classification.”
“Initially we would want to run a fairly rigorous trial where you have a group of individuals who you’re profiling in a controlled and scientific way in terms of their DNA and genomics and microbiome, but then they would be taking something like a Tellspec device and scanning their food,” Dudley says. “It could provide us this rich data on food so we don’t rely on someone saying, I ate this type of food that high fat or low fat or having them make a judgment call on that. Instead we’re going to get the hard quantitative data.”
Hoffmann, meanwhile, hopes researchers will initially combine Tellspec’s more general anonymized data about what people are eating in different locations with data about where certain diseases are most prevalent to better understand disease from a public health perspective. She calls this idea the “food print.”
To do this effectively, however, Tellspec needs to take its beta unit from guinea-pig-size to gerbil-size by redesigning the electronics, battery, and light source. It needs to train Tellspec’s algorithms to identify many more ingredients than gluten and sugar, which means hundreds of thousands of more scans, and eventually, it needs to identify things like pesticides, which will require investment in new spectroscopy technology. Even the relatively crude models of Tellspec that Indiegogo supporters will receive won’t be ready, Hoffmann says, until 2015.
For now, the closest experience to scanning your food, understanding what is in it, and investigating how it could affect your health might be eating lunch with Hoffmann at a Jewish deli (salad: add a scoop of tuna, hold the beans, hold the complimentary bagel).
There’s a link between H. pylori, a gut bacteria, and a predisposition for gastric cancer, she tells me as she pours a pot of tea over ice, and also a link between curcumin [the main active ingredient in turmeric] and reduced H. pylori growth. It’s one of many studies linking food with disease prevention that she has referenced since breakfast.
Researchers might not put these links in such certain terms. “These qualitative associations or simple associations are really not going to be sufficient,” Dudley says. “What you need is real statistics and probabilities to really know the degree to which these things are going to move the dial. It could be that you need to eat 10 pounds of curcumin in order to reduce h. pylori to the extent it would have an actual effect on your gastric cancer risk.”
Tellspec could create data that makes those associations as certain as Hoffmann’s accent makes them sound, if it can bundle a handful of hypothetically possible feats: using spectrometers to scan food, building a food database big enough to include most of what people eat, and figuring out how to add food into the big data map of our health. But if it doesn’t, it won’t be because Hoffmann has given up on the idea. “Go slowly, go slowly,” she says, “We’ll get there.”