By Matt Jones, Senior Product Manager
The amount of patient generated health data (PGHD) being aggregated and analyzed in the marketplace is at an all time high and shows no signs of slowing down. While the amount of data each of us creates on a daily basis, as well as the frequency by which we are creating it, continues to increase, few companies are equipped to make sense of it. How does the amount of sleep a person gets every night affect their blood glucose — or vice versa? How does device usage translate into engagement, or lack thereof, with a remote monitoring program? Can user engagement be predicted?
Validic is uniquely positioned to answer these questions and extend our value past the delivery of PGHD and into predicting adverse or notable events using statistical programming and machine learning. We are excited at the promise of some R&D our data science team has been working on over the past few months.
What we’re hearing from the market
As you may be aware, Validic currently connects to nearly 400 in-home medical devices, consumer wearables, and health apps, which uniquely positions Validic with access to a wide diversity of health data being continuously collected. As Validic continues to add more data sources like continuous glucose monitors and environmental data, the unique nature of our data set will continue to expand.
While we continue to explore our data set and work closely with our customers to solve real problems with data science techniques, there are a few resounding themes we consistently hear:
User Engagement: What good is device data if your users are not consistently wearing and actively using their device? An important success metric of a remote monitoring program is user engagement, but “engagement” is often difficult to define, let alone measure. What if we could tell you when a user is likely to stop engaging with their wearable or remote monitoring device? In a recent predictive machine learning study, we set a goal of anticipating when users would stop tracking activity. We used historical de-identified activity data from thousands of users to train a neural network. Once trained, the neural network was a much better predictor of user activity levels than the baseline approach of using simple statistical correlation. An accuracy rate of 92% was achieved in identifying users that were about to cut off their activity tracking. This method can be extended by curating the definition of an “at-risk” user to align with different use cases.
Program Personalization: Whether it’s an A1C goal or an activity recommendation, our customers must have the ability to customize their communication with their users. It’s no longer useful to send the same message telling your entire population, “Keep it up, you’re doing great!” However, segmenting users based on multiple metrics is tedious and expensive, and manual segmentation can only be based on a handful of inputs, making for weak correlations. Our initial segmentation algorithm used 17 input metrics to analyze users and group them into unique messaging buckets, a task that would require exorbitant effort to perform manually. We are currently working with our customers to iterate on this algorithm for identifying meaningful user segments to power a more personalized engagement strategy.
Validic’s data scientists are working to find the best ways that the latest technology and big data initiatives can help our customers to most effectively and accurately connect with and understand their users, patients, and members. Stay tuned for more updates from our data science team.