A Selfie a Day Keeps the Doctor (a Phone Call) Away
A healthcare revolution was already happening before the pandemic. The old way was the occasional visit to the doctor’s office or hospital, with expensive point measurement, and treatment responses optimized for populations. The new way is continuous measurement of many channels of data about the individual, used proactively for wellness as well as treatment, optimized for the individual. Migraine.ai is on the forefront of this new wave of personalized medicine, using AI to make sense of data that is available every day. Tom got involved with this company when it was focusing on migraine headaches. It was a perfect fit for Humanistic AI philosophy.
Migraines are a big, mostly unsolved problem for millions of people. Although there are a few drugs that can alleviate symptoms, there is no effective prevention or cure. The most important variable is to understand when to intervene and which therapy to apply, based on migraine prediction from triggers. The triggers are sometimes hard to detect, and there is a lot of variability across patients and across time. This is an opportunity for AI technology, which can be pretty good at predicting things from lots of messy data, like the many things that can trigger migraines.
The founder of Migraine.ai, Ozcan Cikmaz, knew that to build these models would require personal data tracked over time. With that in mind, he organized an informal clinical trial of migraine sufferers, asking them about all the factors that the literature says are relevant. Many are subjective, like how much stress you’re feeling. There are a few objective measures that are very helpful: namely vital signs, such as heart rate, respiration, and blood pressure, which are standard measurements when you go to the doctor. If they are taken every day, in combination with subjectively reported symptoms, they can help predict a migraine.
After gathering the data from the early trial, Ozcan found that his team could, in fact, build an AI model that was predictive. In addition to expected correlations, it found subtle combinations of triggers that most people didn’t know were predictive of migraines — especially not the patients themselves. The team discovered that the key to building personalized predictions was daily data.
Ozcan knew that the product was going to need data directly from the patient on a daily basis. Now came the next question: Could he solve the usability problem? It’s both costly and cumbersome to send every patient the instruments to measure vital signs as well as train them in their proper use. How could one get the data from some activity that everyone already could do?
Ozcan knew that the current state of artificial intelligence held out some possibilities for solving this dilemma. He sought out and found a researcher who specializes in inferring vital signs from videos using AI. It turns out that there is enough information to infer vital signs from a short selfie video taken with just about any smartphone, even in poor lighting conditions. Measurements like pulse can be detected by computing the average color of the skin and how it varies over time.
Heartened by this insight, Ozcan recruited a world-class AI team. They started building the app, and began new clinical trials to get the data to refine the models. By early 2020, they were poised to make a breakthrough product in migraine prediction.
Then the pandemic struck. For many startups, COVID-19 has been a major disruption, leading some to stop operations altogether. Ozcan and his team saw COVID as an opportunity to open up a way to serve an even wider population in the field of tele-medicine. It turns out that the key enabling technology in Migraine.ai — the determination of vital signs from videos — is essential for COVID screening. Just as in migraine detection, predicting whether you have COVID is a function of many factors. Among those factors, the objective measurement of vital signs, and the sensitive tracking of those measurements over time, is critical to early detection.
The Big Pivot
So the startup pivoted, while staying true to its core product vision. The company built a COVID screening app called Virologic, which takes daily measurements from selfie videos and other smartphone data such as exercise and sleep when available. Amazingly it even works if you don’t want to show your face; you can put your finger over the camera lens and take a video that works just as well. The company’s AI team has worked night and day since spring of 2020 to refine and optimize these models. Their AI can detect not only the vital signs like heart rate and heart rate variability, but they have developed COVID-specific models to classify the sound of coughs, a key diagnostic of the disease.
When the team made the decision to pivot, some staffers were nervous that they might not succeed. But the bet seems to be paying off. As far as is indicated in published literature and partner companies, as of May 2021 Virologic has the most accurate, comprehensive, and robust of such models in the world. Partner companies are either bundling the Virologic app, or white labelling the technology into their own apps.
The goal for these COVID screening apps is to help people determine whether they are infected prior to getting a biological test. And since Virologic’s screening can be done every day, at home, with no additional equipment, it can be rolled out quickly to very large populations. Studies show that as much as 40% of virus transmission is occurring through people who don’t know they have it. Even a small improvement in the prediction of COVID will make a big difference.
Imagine if this technology could be applied routinely to our video calls, preserving privacy while alerting people to a potential infection. Imagine if this could be given to every essential worker, who is risking their lives to serve the public, and for whom it is essential to isolate themselves if they do get infected. In the “new normal” of post-pandemic telemedicine, this could be part of every video call with your nurse, doctor, or caregiver, giving everyone the best information that could be gleaned from the remote visit, automatically and accurately over time.