InsuLearn Uses Smarter Algorithms to Optimize Insulin Delivery
Arielle Messer and Dorian Goldman leveraged their background in biomedicine and mathematics to develop machine learning software that takes the guesswork out of diabetes care.
Funders, learn how you can become a Health Moonshot Champion and support Health Transformers like Arielle Messer and Dorian Goldman.
When Arielle Messer, a seasoned machine learning scientist, was diagnosed with Type 1 diabetes, she was frustrated with the guesswork required to manage her condition. Despite having access to advanced tools like an insulin pump and CGM device, she found it difficult to make informed decisions about monitoring her insulin delivery given the shifting dynamics that affect glucose levels, such as diet, physical activity, and menstruation. In theory, the algorithms within her pump should have been harvesting her health data in order to effectively monitor insulin delivery. But as a machine learning scientist, she knew these algorithms weren’t doing their job very well. Determined to find a better way, she partnered with Dorian Goldman, a data scientist with a strong background in mathematics, to tackle this issue head-on.
The Challenge
Messer recalls a particular experience where the algorithms behind her insulin pump failed her, causing the pump to continuously deliver insulin over the course of 10 minutes while her blood sugar plummeted. “I almost died,” says Messer. “And that’s an FDA-approved algorithm.”
While the commercial algorithms currently on the market are black boxes that lack transparency, Messer and Goldman were able to infer that the algorithms running Messer’s pump had significant limitations in their methodology. “For example, we found out that the algorithm's forecast appears to rely on the ‘most likely glucose value’ in the next 60 minutes, not taking into account the patients' unique individual risk profile, and appears to lack adequate sensitivity to rapid declines in glucose,” says Goldman.
This severe hypoglycemic incident only compounded the doubts Messer initially had about the technology. “We were already working on this before but at that point, I was like, this is literally insane,” says Messer. “We want to create an algorithm that doesn’t do that, but in the case that it does, someone’s going to be calling you and asking if you’re okay and figuring out the problem. Accountability and transparency is essential for every decision an algorithm makes.”
Messer’s personal frustrations mirror a problem faced by people with diabetes around the world. Twenty five percent of people with diabetes had severe hypoglycemia in the past six months and inadequate pump algorithms are not doing enough to mitigate hypo- and hyperglycemia. Messer and Goldman identified six major problems with the current pump algorithms on the market: no personal adaptability or learning to account for physiological changes, such as illness; a 60-minute forecasting model that lacks responsiveness to rapid changes; lack of auto-correction; patient-guessed insulin parameters, like insulin sensitivity and carb rates; oversimplified linear models to estimate Insulin On Board; and last but not least, the use of “point estimates” of forecasted glucose used to make decisions which is dangerous in the presence of such noisy signals.
The Solution
As Messer was considering how to address this diabetes challenge she turned to Reddit to do some market research. She was intrigued and impressed by some of the open-source products created by the "we are not waiting" community, but realized that the lack of FDA approval and technical expertise required to use these algorithms would ultimately limit the number of people they can reach. Instead she started looking at her data from her pump and CGM device and was soon on her way to building a solution that was more savvy than the existing pump algorithms and took guesswork out of the equation.
When Goldman joined Messer in looking more closely at the problem, he felt like it was something he could help solve. “In my PhD and postdoc I specialized in the calculus of variations and differential equations, with a focus on modeling and optimizing decisions in complex problems that arise in nature,” says Goldman. “I realized that the problem of insulin delivery can be addressed with these tools when augmented with modern machine learning methods, which I had spent the last decade mastering. Given Arielle’s background in biomedicine, we had the exact skills needed to build the best solutions to this problem.”
“Current pumps on the market have the ability to set certain parameters like insulin-to-carb ratio, insulin correction factor, and duration of action,” says Messer, but “all of these things are guesses that the patient or doctor makes.” For Goldman and Messer they wanted to solve this as a math problem and use the data that had “just been sitting there, collecting dust.” They felt that Messer’s data could be activated to set more accurate parameters. Looking closely at the data, they built something that used regression analysis, a statistical method to analyze the relationship between variables and predict the value of one based on others. In this case, they looked at the way Messer’s body reacted to insulin over the past month when she ate carbs. “I could determine that when I eat one carb, I would need to take a certain amount of insulin and I could say that with a level of confidence that can be computed,” says Messer.
This solution – which was used to set initial parameters for a meal – has since grown into a vision for a larger software solution: InsuLearn. This algorithmic product looks at a patient’s history and uses a combination of machine learning and stochastic control, a subset of variational calculus, in order to create a personally-tailored system of insulin delivery. It optimizes glucose homeostasis by predicting dosages based on real-time glucose data and it continually adapts to changes in a patient’s health profile, lifestyle, and treatment response.
How It Works
InsuLearn harvests data from the patient’s pump and CGM device and uses machine learning to analyze that data in order to inform the pump of appropriate insulin dosages. The body’s response is then recorded back into the devices, creating a continuous loop between data, the algorithm, the pump, and the body. The algorithm is designed to adapt to metabolic shifts, including during sickness, and to tailor an ideal treatment model by uncovering consistent patterns in a patient’s longitudinal glucose trends. Ultimately, InsuLearn hopes that people with diabetes never have to think about their pump again.
Where They Are in the Process
Currently Messer and Goldman are developing a prototype based off of the technology that they first built for Messer. It’s a product that will initially only be for doctors as a way to figure out the parameters needed to initialize pump settings and update them over time. “That’s a good place for us to start,” says Messer, who adds that within two years they want to launch a product that can be used directly by patients and integrated with pumps. To reach that goal, they need to develop partnerships with companies that produce pumps and CGM devices. They’re not trying to replace the pump industry but make existing products work more effectively with their smarter software. They’re confident that companies will see value in their product, as many of them don’t have the in-house expertise to create sophisticated software. “This is a business model we see in Europe where software companies have multiple pumps they are associated with,” says Goldman.
Our Take
This product is poised to reduce the mental burden for people with diabetes by increasing their confidence in the devices meant to regulate their health. Through smarter software, InsuLearn can harness one’s health data to make insulin delivery more efficient and ultimately, safer. Arielle Messer and Dorian Goldman are an ideal team to address this challenge, given their complementary skill sets, passion, and experience. Messer is an expert in building machine learning models and has extensive experience developing algorithms for cancer treatment. Goldman has extensive tech leadership experience and brings a deep understanding of mathematics to solving complex technological challenges. Together, we’re confident they have the vision and expertise to transform the way software functions in insulin pumps.
We are excited to have InsuLearn as members of our Type 1 Diabetes Moonshot Community.
Connect with InsuLearn via email
Call for Alzheimer’s Innovation
With support from the Alzheimer’s Drug Discovery Foundation (ADDF) and Gates Ventures, we’ve launched a new global initiative created to develop a collaborative innovation community alongside leading companies, research teams, and stakeholders with a mission to accelerate progress in prevention, diagnosis, and management of Alzheimer’s disease and related dementias. Learn more and apply for an Alzheimer’s Moonshot Fellowship.
Become a Health Moonshot Champion
Health moonshots are fueled by passionate families, foundations, and industry organizations committed to achieving health moonshots. Learn how you can join other champions of the T1D Moonshot or one of our other global health moonshots.
Follow us on social media for daily updates on Health Transformers: X (Twitter), LinkedIn, Facebook, and Instagram.
Published: Apr 4, 2024