Vivienne Ming AIH 2016


I Built a Superpower to Treat My Son's Diabetes

  • April 20th 2016

Vivienne Ming is a theoretical neuroscientist and entrepreneur. She co-founded Socos, which provides services in educational technology and predictive learning analytics, and is a visiting scholar at UC Berkeley’s Redwood Center for Theoretical Neuroscience, where she researches neuroprosthetics. In her free time, she has developed a predictive model of diabetes to better manage her diabetic son’s glucose levels and systems to predict manic episodes in bipolar suffers. She will speak at Spotlight Health 2016.

Four years ago, my family spent our worst Thanksgiving at Oakland Children’s Hospital. It began innocently the preceding Sunday with my son showing signs of the flu. By Wednesday, though, he’d lost 25% of his body weight and was having trouble standing. Squeezing us in before the holiday, our pediatrician broke the news in classic form, asking my wife and I to sit down before telling us it was Type-1 diabetes. We had no family history or experience with diabetes of any kind, but just like that we were parents to a child with a chronic, life-threatening illness.

Insanely trying to balance the mundane and the existential, we scrambled not just to reach the emergency room, but to pick up our daughter from daycare and our pre-ordered Thanksgiving meal from the market. (It turns out it’s possible to be too calm under pressure.) The next 24 hours were the longest of my life as they slowly titrated his blood-glucose levels down to normal range, waking him every 30 minutes with needles and questions. I was left to hold and calm him before grabbing 5 minutes of sleep on a cot next to his bed in the pediatric intensive care unit. The experience was truly awful. But it turns out, the biggest frustration was yet to come.

Our kids have the misfortune of two scientists for parents, and like any good mad scientists, we frequently run experiments on them. As soon as we got home from the hospital we started recording everything. In fact, we crashed Google Docs with our enormous spreadsheets. We recorded everything he ate down the gram, all of his activity levels, his emotional states, every blood glucose level. We built a little cyborg with Fitbits, Basis Bands, wireless insulin pumps (PDM), and continuous glucose monitors (CGM). At our peak, we proudly collected thousands of data points everyday. And in our minds, we thought, “His doctor’s going to love us!”

Not so much, as it turns out. When we brought the pages of data in with us for his first outpatient visit, they were angry. “How dare you waste our time with this? What am I supposed to do with all of this data?” (As an aside, if you ever say anything like this during the course of your workday, someone is already building an AI to turn your job into the modern equivalent of an elevator operator.)

The staff gave us a photocopied form with 15 empty boxes covering morning, afternoon, and evening for 5 days -- "write a blood glucose level into each box” they told us. Completed, his doctor squinted at it for 5 minutes and then offered a treatment plan for the next month. In that moment, fueled by frustration, anxiety, exhaustion, and not a little arrogance, I thought, “You’ve got to be kidding me. I build models of the brain. You’re telling me diabetes is more complex than the brain?”

It was truly crazy that my son was wearing an entire wardrobe of sophisticated technology, yet his pump would be set by hand once every few months based largely on intuition. On that day, I began building Jitterbug, my personal codename for a machine learning system to monitor my son’s diabetes and help us with his treatment.

The first and biggest problem struck almost immediately: all of the data captured by those wearable devices was being sent to some blackbox database. Not one of those companies allowed me to directly access the data that was literally coming out of my son’s body. So I hacked it. In fact, if you go online you can find open source code for almost everything. For example, with a little python code, a couple of USB cables, and an old android phone, I was streaming data off of my son (including fitbits and basis bands) to a server I set up to monitor him in real time.

Jitterbug was originally designed just to do the parameter setting -- carb ratio, basal rate, correction factor -- on his pump, though in real time. I built a Bayesian model to optimally set those parameters right then and there. The easiest way to make that work, though, was to have the system recommend specific doses and then track how his blood glucose levels deviated from its predictions (think expectation maximization for the machine learning nerds reading this).

We built this too, and began piloting it on my son. From this system, he spent 40% less time in a high range. Yet the most amazing and unexpected thing Jitterbug did was actually predict his blood glucose levels as much as an hour into the future. Remember, the system makes these predictions to update the insulin pump. It turned out the predictions themselves could be used to prevent lows. I set up the system to send warnings to my Google Glass (a nearly constant and terrible fashion statement of mine for over two years) so that I knew immediately if my son needed help. In fact, I actually wore Glass to the White House because I was getting a livestream of my son’s glucose levels while he was back in California, on the other side of the country. (The Secret Service was not amused, but I did get to meet the President. He was not amused.) Google: if you want people to wear a fluorescent blue computer on their head, give them a superpower. I would literally have worn Glass anywhere.

Ming's view at the White House while wearing her Google Glass.

I would love to share Jitterbug with the world. As it turns out, however, I’m the fake kind of doctor, and it is illegal for a tool like Jitterbug to give medical advice. This is understandable but also tragic. In the years since we developed Jitterbug, only my son has benefited from it. So we donated all this code to the TidePool Foundation, a nonprofit building open source code for diabetes. It was also wonderful to hear diabetes device-makers announcing this year that their systems were finally making blood glucose level predictions. It only took 4 years and $40 million more than it took me :)

Diabetes has been hard; so many nights, driven from bed by terrible fears. Yet we’ve been so fortunate for those fears to never have come true. In four years we haven’t had an emergency. This horrible experience has turned into something wonderful; an opportunity I wish every parent could experience, to be a superhero for your child. It’s that feeling that you are the one person in the world who can truly change the course of your child's life.

The views and opinions of the author are her own and do not necessarily reflect those of the Aspen Institute.

Thank you for signing up!

Please provide a valid email address.

Please provide a valid email address.
Sign up to receive the latest
updates from Aspen Ideas.