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.
Foundation Model Assessment – Foundation models are popping up everywhere – do you need one for your proprietary image dataset? Get a clear perspective on whether you can benefit from a domain-specific foundation model.

[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, Sam Rusk, co-founder, and Chief AI Officer of EnsoData to talk about optimizing sleep. Sam, welcome to the show.

[0:00:42] SR: Hi, Heather. Yes. Thanks for having me. Happy to be here.

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

[0:00:49] SR: Yes, absolutely. So actually, I’m a co-founder of EnsoData. We started the company out of the University of Wisconsin Madison. We started it right out of our undergraduate and graduate studies. I guess, the background on me, I always had the entrepreneurial bug, per se. I ran like a lawn mowing business when I was in middle and high school. So ended up studying electrical engineering at UW Madison, and was always interested in how the entrepreneurial aspect of how to apply engineering in business and what that meant.

So yes, I ended up kind of getting in touch with my co-founder, Chris Fernandez, and Nick Glattard. I met them in classes like computers in medicine and medical imaging, so kind of the consequence of [inaudible 0:01:30]. That’s where we first learned about sleep, and sleep medicine, and how all these waveforms are being collected diagnostically, and how that could be really a unique application for where we can apply machine learning in healthcare. That was sort of the synthesis in how the EnsoData story began.

[0:01:30] HC: So what does so data do? And why is it important for health care?

[0:01:49] SR: Yes. EnsoData uses AI as one of its primary tools to help automate waveform analysis and spotlight disease processes that have been manifest over different timescales. So waveforms are collected all over the hospital and healthcare delivery system from sleep studies that we work with all the way to EKGs, and cardiovascular monitoring, and different kinds of neurological monitoring. These waveforms are notoriously tricky signals, and that’s where we use machine learning to help process these with generalizability, and trying to help both and diagnostic fronts to identify diseases. But also, where these waveforms collected can help identify therapy opportunities. I think that’s where EnsoData is helping to use AI to bring light to those areas.

[0:02:33] HC: How do you use AI in this context? What kind of inputs, and outputs, and models do you train?

[0:02:39] SR: Yes. Sort of our cornerstone is in the physiologic waveform signal. We use models like convolutional neural networks, and LSTM models to process these one-dimensional time series signals, essentially. Unlike images, and others, kind of a large scale of digital values in waveforms signals. So the kinds of noise and artifacts can be very diverse. That’s where ML helps us see through their noise for a lot of these kinds of signal types that are notoriously tricky. But yes, we’re using a lot of the standard kinds of ML models to take in these essentially time series signals that are collected overnight right now.

[0:03:14] HC: Do you have a couple of examples of maybe the exact things that you’re trying to predict with these models?

[0:03:20] SR: Yes. It’s part of the sleep medicine diagnostic pathway, essentially, the current standard of care is to undergo an in-lab or a home sleep study. And as part of that home sleep study, and in-lab testing, they’ll collect pulse oximetry, which is the fingertip monitor, they’ll collect breathing rate, both through the flow, and kind of the nasal pressure. But then, also, around the chest to understand the effort, and in the breathing overnight. There’s also neurological signals, so electroencephalogram, EEG signals that are collected.

All of these are inputs that are collected diagnostically. Then, from there, our software helps to analyze things like, which stage of sleep you’re in. So of the five stages of sleep, we have wake, and one, and two, and three in REM. Those are categorized by our product, and usually by sleep conditions. We also identify respiratory events into saturation events, where patients will stop breathing, or they will desaturate over the night. If there are also arousal and like movement events, which are kind of explained. But essentially, when a patient will wake up, that’s part of the standard diagnostic process, they’ll mark an arousal. And then when a leg moves, the same there. There’s kind of a host of different events, essentially is what they’re called in sleep medicine, that our product helps to detect.

[0:04:31] HC: In order to train models like these, you would have someone go through and annotate these different types of events. Is that how you base these models on?

[0:04:38] SR: Yes, exactly. We have large data set that we’re always growing, collecting from partners. As part of that, we often receive these annotated datasets, and we also facilitate some of our own data annotation through our internal waveform viewing System. But as part of that, yes, we’re learning to validate, and verify our models on datasets that are annotated, both from rural settings, and also for more of a research setting as well.

[0:05:03] HC: What kinds of challenges do you encounter and working with, and in particular training models on waveform data?

[0:05:08] SR: Yes. There are – as I said earlier, waveform physiologic signals are notoriously noisy, just with kind of an undefined level of variability that can happen. A lot of the challenges that we do are kind of learning to generalize our models across different sensor types and different signal modalities. That’s where I think we spend a lot of time is, taking a look not at the overall performance, but sort of the outlier performance of what are the examples, or what are the situations that give us the most challenge, and how can we look at, maybe have a lower overall performance, if we can capture a lower variability on that performance. I think that’s where the waveform, and these artifacts can offer a lot of challenges to making sure that you’re not leaving a certain kind of data as an outlier for diagnostic performance.

[0:05:54] HC: Generalizability is a challenge, and a lot of different application areas, whether it’s from a different facility, a different patient population. How do you tackle that kind of challenge?

[0:06:04] SR: Yes. We do a lot of verification and validation testing. A lot of it is really continuing to expand, and build upon the diversity of the data set that we’re using for testing. So, historically, our strategy has ultimately been to validate. and verify on a subset of devices, or different kinds of data that we have. Then, as we collect more, we introduce those into our testing. and try to understand performance. If there are deviations, then look to improve those.

It is sort of an iterative process for us right now. I think, ultimately, we hope to use more complex, more sophisticated generalizability, and generalizable machine learning methods in the future. But yes, it’s sort of this process where there’s always new devices, there’s always new sources of data, and we have set up systems to help measure those over time.

[0:06:51] HC: It sounds like a very data-centric process. Are there any more model-centric approaches that you use to tackle generalizability? Or is it mainly data-focused at this point?

[0:07:01] SR: That’s a great question. I think it is very largely data-driven on sort of an observability. But we’ve implemented and experimented with many different kinds of generalizability approaches for the machine learning modeling, using different kinds of bias losses. That is, you train your models, you’re minimizing the loss over the different kinds of groups that you need to make sure that the models don’t have deviation over. I think we’ve gotten all the way down to the literal training process trying to tell the model when there is variability.

I think that’s – we’ve tried a lot of methods in line with that, and I think we hope to continue to explore that. As we see the future of machine learning, I think that as specifically applies to healthcare problems is a huge important topic that people are talking about is generalizability. Specifically, what are the dimensions of generalizability across different gender, sex classification, ages, areas in the country, areas in the world. I think that the young area and machine learning in healthcare still.

[0:07:58] HC: You co-founded EnsoData nine years ago. How did you think about the role of machine learning in the earlier days? Perhaps, how has it changed?

[0:08:07] SR: I think the methodology behind machine learning and deep learning has obviously evolved quite a bit over the last – what’s now, seven or eight years of the EnsoData journey. I think, interestingly, a lot of the problems that we’re looking to tackle are still relatively similar to when we started the company. Not that we don’t understand them better, or that they’re more developed. But I think the rate of technological innovation on like a clinical side is a little slower than where it is, especially on a deep learning, and more academic field.

A lot of the models that we’re deploying are not necessarily the most cutting-edge, day-to-day for the machine learning world. But we’re continuing to use the tools to help push the clinical opportunities forward. That is ultimately a big problem that we also face, aside from just the core technology that we’re building. So yes, I think machine learning itself has obviously come a long way. But we’re still working on the same problems, and we’re five or six years in, so we’ve come a long way, and I hope we can catch up to where a lot of the modern-day papers and publications are.

[0:09:08] HC: How does your team plan and develop a new machine learning product or feature? In particular, what are some of the early actions you take in this process?

[0:09:17] SR: Yes. As a company, we do – our core product, Enso Sleep is a software-only medical device. As part of that, we operate with a quality management system, kind of fall under regulation with hosting and delivering a medical device. As part of that, we have a pretty sophisticated feature, and kind of roadmap process that we work with in our engineering. Basically, it starts with an idea that turns more into a defined project or feature. Then, our development team picks it up, and scopes it from there, and then completes development, and then we release it.

We kind of have a pretty mature process for taking feature ideas and moving them from the top of the funnel on product management all the way to testing and releasing those. That’s something that we spent a lot of time on as a company, and I hope that we can continue to mature that as we see that it’s a way for us to make and provide safe products that we can move with high velocity. But yes, if we kind of have this really built out pipeline for moving features into production.

[0:10:14] HC: Could you elaborate on some of that process? What steps and which team members do they involve? Do they do a lot of research beforehand, or do they just dive into playing with data and machine learning?

[0:10:27] SR: Yes, no. There’s a lot of research. I think, fundamentally, it’s in a few steps. The first step is the idea, and sort of the definition step, where product management plays a huge role. I think that’s an array of that company that we’ve built up over the last couple of years, more so than previous. That’s really, yes, is this the right thing to build? Is this something that is important to our customers, and will actually make a difference in our customer’s day-to-day, and their service that they provide to patients. That’s really majority of the – hopefully, the filtering out of ideas, and what it is that we build on the long-term roadmap.

Then, from there, the next phase is what we call scoping, which is where we go try to understand on a technological level, what that idea, and what it would take to build this feature. That is an important step so that we can help to plan, and deliver our product roadmap on a long-term timescale in a sort of predictable manner. Then from scoping, it moves into development and testing. That’s sort of the classic engineering processes of making something, and then testing it appropriately. But those are sort of the pipeline for how we get in. The cool thing that we do is, I think anyone in our company can create an idea at the top of the play. Anyone has the power to say, “Hey, this might be a good thing that we want to give to our customers.” Then from there, product management is able to take this big pool of ideas and ultimately get down to with the highest impact things we can do today.

[0:11:43] HC: How do you ensure that the technology your team develops will fit in with the clinical workflow and provide the right kind of assistance to doctors and to patients?

[0:11:51] SR: Yes. I think, as I said, the product manager is doing a lot of the early leg lifting. But we have, I believe seven different RPSGTs, which is sort of the clinical sleep credential as a sleep clinician. We have seven team members that are team that do a lot of the testing and really get into the product before we release it, and help to give feedback, and help the guide even at the very last [inaudible 0:12:10] show. We’re continuing to build in that feedback loop of our internal clinician team, and our close external stakeholders who are going to use the product so that we can kind of release things, and create a tight feedback to provide the best experience for our customers.

[0:12:25] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[0:12:30] SR: Yes. I think there’s still – in healthcare, there’s still a massive opportunity to build machine learning and AI-powered solutions that can really change the game for healthcare delivery. It can be difficult to find, to set up an entire business model around a lot of these applications that are needed. I think being thoughtful about the economics of your business, and how the go-to-market strategy, and really sort of the business side, I think there’s a lot of hype around AI, obviously, and there’s definitely a lot of things to do in healthcare. But we spend a lot of our time solving that, not necessarily the machine learning performance side of the problem, but more of, how do we get this into the clinician’s hands in a way that makes sense for everyone.

I think that’s where – there’s still just a lot of barriers to getting new ML and AI solutions into healthcare frontlines right now. That’s where I would probably recommend the most, is to focus on the entire process of what it means to deliver AI in healthcare and in other fields as well.

[0:13:25] HC: Finally, where do you see the impact of EnsoData in three to five years?

[0:13:28] SR: Yes. Our vision is to continue to bridge the gap between diagnostic and therapeutic pathways, especially in sleep. There’s pretty significant fragmentation between when a patient shows up to get tested for a condition, and how they’re treated, and their long-term management. We hope to use machine learning for every single step in that process, know, from the start to diagnose, and then to help patients find the right therapy that’s most effective for them. Then ultimately, to help manage that if it’s a long-term kind of therapy.

Our vision is to continue to grow the gap. We want to be a research leader in machine learning. We want to deliver products that change the game, we invest heavily in research, and we are active in the community publishing, and engaging in the research community in sleep. I think I hope to continue to see that part of our company grow. I think we’re just kind of scratching the surface for ultimately machine learning, and these kinds of waveform, and other data can do to improve patient outcomes.

[0:14:26] HC: This has been great, Sam. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[0:14:34] SR: Ensodata.com, you can check us out there. We’ve got a ton of different content, and with education about what we do in our vision. That’s where –

[0:14:42] HC: Perfect. Thanks for joining me today.

[0:14:45] SR: Yes. Thanks for having me. It’s been fun to talk about our company.

[0:14:48] HC: All right, everyone. Thanks for listening. I’m Heather Couture, and I hope you’ll join me again next time for Impact AI.

[OUTRO]

[0:14:58] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe, and share 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]