LinkedIn – Connect with Heather.
Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.
Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.

[INTRODUCTION]

[00:00:03] HC: Welcome to Impact AI, brought to you by Pixel Scientia Labs. I’m your host, Heather Couture. On this podcast, I interview innovators and entrepreneurs about building a mission-driven, machine learning-powered company. If you like what you hear, please subscribe to my newsletter to be notified about new episodes. Plus, follow the latest research in computer vision for people in planetary health. You can sign up at pixelscientia.com/newsletter.

[EPISODE]

[0:00:33] HC: Today, I’m joined by guest, Bill Tancer, Co-founder and Chief Data Scientist of Signos, to talk about metabolic health. Bill, welcome to the show.

[0:00:42] BT: Thanks, Heather. It’s pleasure to be here.

[0:00:45] HC: Bill, could you share a bit about your background and how that led you to create Signos?

[0:00:49] BT: Yes, absolutely. I got involved in data science, let’s see, going back to around 2004 when I joined a small Australian startup called Hitwise. Back then, we were measuring what people were doing online. So I got a sample in the US for 10 million Internet users and 25 million worldwide. I was responsible for calling through all of that data and finding tactical uses for it.

In the process of doing so, I became a data evangelist. I found these really fascinating stories within the data set and then proceeded to write about it. I wrote a column for Time magazine called ‘The Science of Search’. Published a book called Click: What Millions Do Online and Why It Matters. Did that for quite a few years, and then we were acquired by Experian. So then, my role expanded. I had access not just to internet data, but all the data that Experian had access to, and continued my role for about 12 years.

After that – I’d say, during that timeframe, I also was advising a good friend of mine, Sharam Fouladgar-Mercer on one of his startups, and he was starting up a new venture. This is back in 2019. We had dinner in San Francisco to talk about this idea of using CGMs or continuous glucose monitors to help people lose weight. I thought it was a fascinating idea.

Coincidentally, at the same time, and I hope you don’t mind if I share a personal anecdote. But about the same time, I had just been in to see my doctor for my yearly physical. He had come into the room with the labs from that visit, and he said, “You’re butting up against metabolic syndrome. Your glucose numbers are too high, you’re at 99, which is right below, one point below being pre-diabetic. Your triglycerides are too high, your cholesterol is too high, you need to lose some weight.” He said, “You need to do something. You’re going to have a bunch of health issues.”

My question to him was, “What should I do, doc?” His response was, “Well, I don’t know. Lose weight.” That’s all he had in terms of advice for me. It’s kind of ironic that at almost the exact same time I’m sitting over dinner with Sharam talking about this idea of using CGM data to help people lose weight, I had a personal interest. At first, decided I’d be an advisor for the company. But within a few days of using the CGM and using the Signos platform, I was convinced that I wanted to play a more active role, and I became chief data scientist at Signos, then co-founded with Sharam, Pierre, and Dr. William Dixon, our medical co-founder.

[0:03:39] HC: What does Signos do, and why is this so important for health care?

[0:03:42] BT: Yes, sure. Signos translates an individual’s metabolic signals and those metabolic signals that we deal with our glucose readings using the best in class CGM out there, which is Dexcom CGM. We take those signals, and we turn them into timely recommendations for what you should eat, when you should exercise, how much exercise you should do, all to achieve better health and weight loss.

When a member joins Signos, they put on the CGM, they log into our app, and they log in what they eat. That allows us to look at each individual’s glucose readings for specific foods. Once we’ve calibrated, Signos provides data that’s personalized to the individual, personalized nutritional data that suggests what you should be eating, how much you should be eating, when you should exercise to mitigate a glucose spike, and things like that to keep your glucose within the optimal range. By doing so, reducing a lot of your glycemic variation. In terms of why this is so important, the numbers are quite alarming. I learned these as I was joining Signos. But if we look at the US, there are 96 million Americans that are pre-diabetic. You add type two diabetes on top of that, and there’s 134 million in the US that have problems with glycemic dysregulation. That’s about 40% of the US population.

Along with diabetes as its own health risk, having that dysregulation of your glucose can lead to a host of other medical problems, such as cardiovascular disease, strokes, Alzheimer’s, just to name a few. We think this is such an important goal for the company to help people reduce their glycemic variability. In doing so, get their glucose down from where it is, and help to get themselves into more of a normal range. And in the process, lose weight and become healthier.

[0:05:49] HC: What role does machine learning play in this technology?

[0:05:52] BT: Yes, sure. We use ML in a number of different ways. The two primary ways are based on what you eat. We use machine learning algorithms to predict your glycemic response to that food. Then the second thing that we do is that, based on that glycemic response, we estimate the amount of physical activity you would need to mitigate that glycemic spike that you might be experiencing from a specific meal. A lot goes into that algorithm. Essentially, we’re looking at how you respond to specific foods, we calibrate during a period of time. Then, based on future entries to your food log, we can make those predictions.

We can also see when you do move, when you exercise, what effect that movement has on mitigating that response, and then figure out what amount of physical exercise both in intensity and time you might need to mitigate a response. What our members experience is, just those two pieces of information, some guidance on what it is you should eat, and when you should move are enough to help them to start move the needle down in terms of their average daily glucose, or how their glucose is varying throughout the day. They experience two things, they experience weight loss, but they also experience overall wellness in terms of a number of different factors.

[0:07:19] HC: In order to train machine learning algorithms using the CGM data, you need to gather it. In your case, do you need to annotate it as well, and how do you go about doing that?

[0:07:30] BT: Well, gathering the data is relatively simple. So we’re taking a signal from the CGM and ingesting that into our own platform. Dexcom supplies glucose reading every five minutes. That reading is actually the glucose, that’s an interstitial fluid. So it’s taking that number, it’s interpolating it into blood glucose. We take that, and we ingest it into our system. It’s really as simple as that, and then watching that time series of data over time to see how it varies. Now, the tricky part is then getting our members to log their food, and a lot of people do, but it can be a challenge to get people to log food. We then associate a particular meal with a glycemic response, which can happen usually 20 to 30 minutes after a meal, it can last as long as two plus hours after a specific meal. That’s the first source of data, coming in from the CGM.

There’s other sources of data that we use, a lot of it is member-supplied data. I mentioned nutrition logs, but we also ask our members to weigh themselves. We asked them to log exercise, sleep, and hydration. Also, members do opt-in to provide some of those health metrics. In addition to those, things like HRV, which can help us estimate things like stress, as well as some of their heart rate data that gives us a better idea of the intensity of exercise.

[0:08:58] HC: What kinds of challenges do you encounter in working with CGM data, in particular, in training models based off of it?

[0:09:04] BT: The biggest challenge I would say is probably that number supplied data, so correlating action and glucose reaction. Not all members log religiously, and sometimes, the accuracy can vary in terms of how people log. This isn’t a challenge that we face in Signos. This is an industry-wide challenge that we have in terms of people logging food. That being said, we try to mitigate that challenge by allowing people to do things such as just scan the barcodes of packaged foods. They can even log meals as text if they don’t have an idea.

I know personally, my biggest challenge, when I’m faced with food logging is when I go to a restaurant. I was just at a Thai restaurant the other night. I’m probably not going to go back into the kitchen with a scale and ask the chef if I could weigh everything that he’s putting into the meal, or even ask him all of the ingredients that exist in that specific meal. Sometimes, when you’re faced with those big challenges like that, and I think this happens to some of our members, they may not respond or may not log a specific meal, so that’s a challenge.

That being said, we can still see the spike, and we have used them out in another way. That is in looking at a possible meal based on what we’ve seen in members’ history. We can actually flag a spike, and then intercede and ask somebody if they ate something at that period of time, and then inquire what they might have eaten. Even if we get text, that helps us train the ML model. I would say that’s the biggest challenge.

The other challenge is, is that, there’s a lot of data that we can use, and this is looking more out to the future. Not all of those data sources are standardized, and not all of them are mature. Using HRV to estimate stress is a great example of that. There are a number of different sources that gather HRV data. Be it the Apple Watch, and its HRV estimations, or what you might get from a Whoop or an Oura. I think, over time, we’re going to see those data sources become much more mature and reliable. That really will put our algorithm in turbo and we can start to incorporate a lot of that data, and make some recommendations based on things like stress and sleep.

[0:11:21] HC: How do you ensure that your models continue to perform well over time? Maybe if the user changes their diet, or their health changes over time, or external factors, does that influence model performance and how do you mitigate that?

[0:11:34] BT: Yeah, absolutely, it does and things change. To start off with, we have this advantage of having this closed-loop system. So we monitor our members’ glucose, we gather a variety or calculate a variety of different metrics from time and range to average daily glucose, estimated fasting glucose. We calculate GRV, which is our form or version of a coefficient of variation. All of those in addition to what the desired outcomes are. The first for a lot of our members is weight loss, so we’re actually getting weight loss data. Then, just increasing an individual’s glycemic control would be another outcome that we’re monitoring. So that closed-loop system helps us evaluate how well the model is doing, and predicting how well the member will achieve their outcomes.

In that sense, that’s the great thing about our data, is we’ve got the inputs, and we also have the desired outcomes, which helps us monitor over time. You did bring up a really, really good point. I’ve been using our platform for over three years now. It’s interesting that as I’ve gotten healthier, as I’ve got a lot of my own metrics within a range that are considered healthy, my body’s changed. Some things that were true in the past are not true now. I respond to things differently. Even my body’s response to a specific food can change. While those are normally challenging, given that we have that closed loop, that we’re always looking to the desired outcome be it continued weight loss. I don’t need that anymore. But, maybe now, it’s just maintaining my glucose within a given range. The model is actually recognizing that and adapting.

[0:13:20] HC: Thinking more broadly about what you’re tackling at Signos, how do you measure the impact of this technology?

[0:13:26] BT: Well, there’s that. We’re measuring those outcomes, which I think are key. If we can help all of our members get their glucose under control, move them from being pre-diabetic into normal range, that’s a victory for us, and we can see it within our metrics. But I have to say, in addition to that, it’s looking at some of the qualitative responses, and the anecdotal feedback that we get from our members. Be it in our private Facebook group, or what they’ve been telling our customer support team.

Things like their hemoglobin A1C has gone from 10 to 5.6 while using our system. Or the fact that they’ve got improved fitness, or their quality of life has improved. That is incredibly moving. That’s what gets me up in the morning, is hearing those anecdotes, but it’s that in conjunction with looking at our own data and how our members are improving in terms of their wellness that really tells us we’re having a measurable impact.

[0:14:32] HC: AI has been in the headlines a lot lately with generative models and particularly large language models like ChatGPT. How do the latest advancements like this influence what you’re working on or does it influence what you’re working on?

[0:14:46] BT: It doesn’t really to date. Not to say that there’s not going to be a place for generative AI going forward, and I can imagine a few different places where it could fit within the business that we’re in. To date though, the use of AI is limited primarily to our ML algorithms.

[0:15:05] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[0:15:10] BT: That’s a great question. I’ve been monitoring so many different AI companies, especially in healthcare and beyond, actually. I think one of the things that has been so important for us, and what has really helped us as a team is having that big problem to solve and solving that specific problem as part of our mission, and always being focused on solving that specific problem. I say this because it is so easy in the space of all the things you can do with AI to end up in a space where you’ve got a solution that’s searching for a problem to solve. The antidote to finding yourself in that situation is always returning back to your mission.

For Signos, we have 40% of the US population struggling with glycemic control. That is on our minds every day, so we’re always thinking about the technology and how it can solve that specific problem. If we were to not have that mission in mind, I think it could be very possible we go off doing all sorts of different solutions that really don’t impact what we set out to do, and having the discipline to always return to mission. Figuring out how can we make this algorithm better? How can we help our members achieve this glycemic control and achieve wellness? It’s that. It’s having that front and center because it really is a huge sandbox you can play in with this technology. Keeping yourself on-mission is critical.

[0:16:47] HC: I think that’s a commonality amongst many mission driven startups. They’re succeeding because they’re focused on what they’re trying to achieve. If AI is a solution, that’s great. If something else is a solution, that’s fine too. But staying focused on what they’re trying to achieve, and then just finding the right technology to solve it get some further ahead.

[0:17:07] BT: Yes, absolutely.

[0:17:09] HC: Finally, where do you see the impact of Signos in three to five years?

[0:17:14] BT: Unfortunately, this problem is not going away. In fact, I just saw something that was published in Lancet. This was this summer in 2023. Estimating that, if we look worldwide, the projections are, there’s going to be about a billion people that have type two diabetes by 2050. The projections in the industry are that, this problem is just going to be getting worse. We’ve got our work cut out for us, but I think that we could make some headway in lowering that number by providing data-driven, ML-driven algorithms that help people make healthy decisions on a daily basis. One of the things I didn’t mention is back, last year 2022, we launched the first of its kind study, it’s a 50,000-person study on the effects of using a CGM, with a mobile app to help people with their health outcomes.

These are specifically non-diabetic individuals, healthy individuals. In the next few years, probably in that three to five timeframe, we’re going to have the results of those findings. I am very hopeful that we’re going to be able to provide some amazing insight on how people can really maintain or achieve metabolic health.

[0:18:35] HC: This has been great, Bill. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[0:18:42] BT: Yes, sure, Heather. They can find us at signos.com. Our social handle is Signos Health. I also discuss these topics a lot on our own podcast, which is called Body Signals, which you can find on Apple podcasts, Spotify, or wherever you consume your podcasts.

[0:19:02] HC: Perfect. I’ll link all of that in the show notes. Thank you for joining me today.

[0:19:06] BT: Thanks, Heather. It’s been such a pleasure.

[0:19:08] HC: All right, everyone. Thanks for listening. I’m Heather Couture and I hope you join me again next time for Impact AI.

[OUTRO]

[0:19:19] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe, and share it with a friend. If you’d like to learn more about computer vision applications for people in planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]