blog

Leveraging Applied Health Signals to Reinvent and Scale Digital Coaching

Anmol Madan

03.25.20

Want to learn more about Livongo?

A solution for managing efficient care delivery during, and beyond, COVID-19

Anmol Madan, Chief Data Scientist, Livongo

From its earliest days, medicine has centered on human interaction — particularly the intensive and personal “encounter” between physician and patient. Over time, as healthcare systems have grown to meet demand and complexity, the circle of care expanded to include an ever-widening array of roles — nurses, medical assistants, care managers — with workflows directing and delegating care to the right provider at the right time. Now, with the coronavirus pandemic putting unprecedented strain on hospitals and physician practices, we’re seeing the critical role that companies with remote monitoring capabilities can and must play to help proactively track a person’s health in real-time and at scale, delivering 24x7 telehealth access to certified health coaches, and the targeted guidance people need to stay informed, healthy, and safe. (My colleague and Livongo President, Dr. Jennifer Schneider, addressed Livongo’s coronavirus response here last week.)

In recent years, as chronic condition management has become a critical national priority, and smart, connected technology enables new levels of interaction, health coaching has come to play a pivotal role in facilitating long-term behavior change. In 2018, health coaching was a $6 billion service market with a strong growth outlook, with 81% of healthcare organizations incorporating coaching into their care delivery models. The market is recognizing that the combination of digital guidance and health coaching can not only help scale access to care but can offer consumers a more personalized and in-home care experience. When machine learning is added to the mix, the result is a dynamic learning system that can predict, and personalize solutions to the individual and their needs.

At Livongo, we have grown to serve over 220,000 Members, most of whom are at-risk for serious health complications, and more than 800 clients, including over 30 percent of Fortune 500 companies, four of the top seven health plans, and partner with two of the nation’s largest pharmacy benefit managers. We deliver multiple coaching modalities across our diabetes, diabetes prevention, hypertension, weight management, and behavioral health solutions, including digital coaching and Health Nudges™, 24/7 alert support, chat- and text message-based coaching within the Livongo mobile app and scheduled 1:1 coaching sessions.

To keep pace with this momentum and support our Members, we are continuously scaling our digital coaching capabilities and refining our approach to integrating high-touch coaching with AI-informed digital guidance. This means deepening our understanding of our Members and their needs; determining when coaching resources are in demand; and treating every interaction as an opportunity to learn from Members and improve our support and the overall experience to drive positive behavior change.

Understanding Our Members

While much of healthcare is still mired in one-size-fits-all service models, deep personalization through machine learning has become standard operating procedure in retail, media, and other industries. Whether it’s Netflix suggesting what to binge-watch next or Sephora recommending products based on skin type and beauty preferences, leading retailers are using rich consumer data to help remove friction and delight consumers.

At Livongo, we have diabetes response specialists available 24/7 who are alerted to contact a Member if their blood glucose (BG) readings go out of range. To consistently meet or exceed our response time, which has averaged 20 seconds, we are continuously identifying and predicting when Members are likely to place demands on this service. An analysis of historical Livongo platform data representing more than 100,000 Members (Figure 1) tells us that people who typically trigger these alerts have a higher mean BG value, greater variance in BG levels, and more checks on the Livongo platform using their BG meter. By matching these characteristics against our entire Member base and tying in the dynamics of our business (e.g., enrollment growth, deals signed), we can begin to anticipate Member needs and speed response time. As we continue to ingest more signals from more sources and devices —including the leading smartwatches and CGMs — they enrich our understanding of the best way to support our members via digital and human coaching.

Figure 1- Image

FIGURE 1

Predicting Member Demand to Guide Supply

Rapid response health coaching is a resource intensive service. Understanding who will require more clinical or coaching support is a good starting point for scaling services. From the moment we enroll Members, we incorporate rich data from pharmacy claims, electronic health records, and claims data and combine that with health signals collected from our proprietary devices to build out detailed Member profiles. The next step is to predict and prepare for when and at what level those services will be in highest demand.

Supply-demand forecasting is well understood in industries outside of healthcare. Power grids are constantly adjusting power generation based on a wide range of load-predicting factors, including weather, price, and consumer behavior. Rideshare companies like Uber track a range of variables — weather, sporting events, holiday traffic — and both boost supply and impose surge pricing to manage flow. Healthcare remains notoriously behind the curve on this kind of forecasting, leading to many of the challenges we are experiencing today including lack of resources, epic wait times, and widespread frustration — even while the same facilities often operate below capacity.

At Livongo, we’re using machine learning to solve this challenge by building a dynamic and responsive system that can accurately anticipate and plan coaching resources to meet Member demand. As illustrated in Figure 2, we are using historical data on weekly Member alerts to our diabetes response specialists to train and refine a forecasting model that can accurately predict future demand. The result for our coaching system is the equivalent of a physician practice that always knows how many people will walk in and with what needs, has detailed data on their progress between visits, and can therefore staff providers and personalize care accordingly.

Figure 2- Updated

FIGURE 2 (Note: Figure uses previous prediction data)

Learning From Every Interaction, Every Signal

A key tenet of Livongo’s Applied Health Signals strategy is to consider every interaction with our Members as an opportunity to know them better and more deeply personalize how we support their care journey. This means knowing when and why they trigger alerts. It also means listening closely to what they say when they seek support.

Thanks to our partnership with Amazon’s AWS Transcribe team, we’re now using natural language processing to capture (in de-identified, aggregate form) every coaching session, transcribe and analyze it for qualitative clues to member concerns and sentiments. When we look at the word clouds and topic clusters that emerge (Figure 3), we can begin to gain fresh insights into member conversations and concerns. For example, when Members speak with coaches do their conversations cluster around diet and exercise? Or do they talk about managing stress or other pressures that might result from or exacerbate their condition? This kind of deep sentiment analysis will allow us, over time, to better understand and parse the complex relationships between diabetes, stress and self-management so we can move from treating isolated conditions and instead address the whole person.

Figure 3- Updated

FIGURE 3 (Note: NLP Data pulled 02/20)

Both within and outside of healthcare we are witnessing the rapid evolution of human-machine interaction. Last year, Google announced and began rolling out BERT (Bidirectional Encoder Representations from Transformers), a neural-network-based technique for NLP pre-training that moved computers that much closer to understanding language the way humans do. That release marked “one of the biggest leaps forward in the history of Search,” as Google’s VP of Search, Pandu Nayak, put it. One advantage of these transformer advances is that they enable “transfer learning,” wherein pre-trained models can be applied to a new but related language task. This is helpful as many healthcare applications are working with limited datasets.

Over time, the gap between digital and human understanding, conversation and coaching will continue to narrow as AI gets smarter and more scalable. Machines will never fully replace the judgement or empathy that another person — whether doctor or coach — can offer. But with COVID-19 challenging every aspect of our healthcare system, it is apparent that we must offer remote solutions that deliver high care at scale. And when people are interacting with their physician or health coach, machines can and will continue to make those moments of human interaction ever more informed, more personalized, and more effective at improving outcomes.

Anmol Madan is Chief Data Scientist at Livongo where he is responsible for machine learning and data science across all of our products and services.

All great things are done by teams. This forecasting project was led by Livongo data scientists RJ Ellis, PhD and Jesse Bridgewater, PhD. The work in natural language processing was led by Amin Mousavi, PhD and Karthik Kappaganthu, PhD. All projects were supported by Whitney Mirro, Rebecca Mitchell, MD, Anisha Narula and others.

Further reading

blog

Partnering with Kaiser Permanente to Support Mental Health Through COVID-19 and Beyond

From Dr. Julia Hoffman - 04.02.20

blog

The Coronavirus Stimulus -- A Step in the Right Direction

From Leslie Krigstein - 03.30.20