Omada Health's Journey with Artificial Intelligence

How Data Science and AI learnings surprised us, taught us, and propelled us forward

The buzz around AI, and more specifically, Generative AI (GenAI), has led to questions about applications and use cases in healthcare, and where the intersection of human-led care and evolving technology will shape the future of virtual care. For Omada, the topic of AI isn’t new. We’ve been learning about AI for over a decade, devoting time and resources to form our points-of-view, and explore the role this powerful technology can play to empower compassionate human care.

Since Omada’s founding in 2011, the broader mission has always been tackling chronic disease, which remains a serious problem in the United States. In all, 6 out of 10 Americans currently have a chronic condition, and people with multiple chronic conditions drive 71% of U.S. healthcare spend. Diabetes alone affects over 38 million Americans, approximately 47% of whom don’t have it under control. 

With deep experience and longevity in virtual health, we believe Omada is well-positioned to make a meaningful impact in the fight against chronic disease. We believe we’re equally well-positioned to leverage data to transform personalization, benefit outcomes, and deliver richer, more meaningful member experiences. With over 1 million all-time members, we have a robust dataset on chronic disease care – at a moment in technological history where data is the currency, not just for Machine Learning Models, but for Generative AI. 

Over the years, we’ve journeyed through different forms of computational intelligence and developed learnings on the challenges, revelations, and philosophies behind AI in healthcare. With that in mind, we’d like to shed some light in the event our stories are useful to others.

“Freeform Coaching” vs. Data-Science-Driven Coaching

The tool in the first years of Omada was primarily a messaging input and a weight chart. That was it. The number of members I could support in that world was much lower than it is today.”

Without systems in place and data to draw on, scaling the work of health coaches to impact chronic disease outcomes on a population health level would prove difficult, and Omada believed that getting care teams to work in concert with technology would be vital. This belief sparked the creation of an official data science program circa 2015. 

With this data science program, our engineers implemented a data-led coaching model with a set of rules and heuristics. This system provided certain suggestions for coaches, but only concerning a narrow set of tasks––i.e. basic reminders, or what to message and when. Outside of coaches’ normal day-to-day work, the system required that coaches complete certain tasks, one at a time, and managed coaches’ performance in part based on adherence to each task. 

This quickly created tension between data-science-informed coaching and the freeform style coaches had grown accustomed to in years prior, recalled Omada Senior Director of Health Coaching, Devin Ellsworth. Without knowing when coaches complete their recommended tasks, it would be harder to isolate and understand their impact; that became the message relayed to coaches. “The point is to challenge relying only on your intuition,” Ellsworth remembered telling coaches. He and others recall the tension between encouraging coaches to work completely autonomously and embracing support from data science, which felt to some coaches like serving the algorithm. 

The Aftermath and Lessons Learned

Over roughly a three-year run (2015-2018), leadership observed modest improvements in outcomes, including for weight loss, according to Ryan Quan, Omada’s current Director of Data Science. “That mobilized a lot of resources toward that model during that time,” he said. 

However, the friction between coaches’ desire for a more freeform approach and increasing suggestions from data science was inefficient operationally and prevented the data science program from scaling. A lack of exploration and innovation also plagued this period because the platform only recommended one task at a time for coaches, Quan said. “We were only able to evaluate single ideas at a time,” he elaborated. “We weren’t getting enough diversity of hypotheses to analyze. We were very good at optimizing and exploiting what we found, but not as good at searching for different opportunities.” 

We weren’t getting enough diversity of hypotheses to analyze. We were very good at optimizing and exploiting what we found, but not as good at searching for different opportunities.”

Once Omada decided to expand and offer multiple products in the market, including Omada for Diabetes and Omada for Hypertension, Omada divested from the original iteration of the data science program and its analyses and development. “The system stopped scaling, so we started to figure out what was next,” Quan said. 

Omada also valued the human component of coaching programs, and Quan, Ellsworth and others believed that striking the right balance of AI and human-led coaching should drive the coaching service’s entire structure. Compelling data also suggested that the humanistic aspects of coaching appeal to members. In a 2019 Omada survey across a variety of demographics, 62% of consumers surveyed responded that they wanted a human health coach for guidance and motivation if one was offered to them, while 84% wanted a human involved in their care no matter how advanced technology becomes. With this understanding and more data available as our program grew, Omada transitioned to a model that more greatly emphasized how data-led insights, informed by rules and statistical machine learning, could help coaches scale their efforts while still offering them autonomy. 

Omada redesigned the platform to include a participant-centered inbox that could surface multiple data-informed coaching tasks for the same member, increasing the level of autonomy for coaches. A coach could determine, of three or four suggested tasks, such as timely diet or exercise interventions informed by predictive analysis or forecasting trends, which is most relevant and appropriate for their member. Coaches were also offered a clear option to choose "none of the above," if none of the data-informed tasks fit the moment, a critical step in emphasizing coach autonomy and allowing our care teams to follow our members’ lead first. The participant-centered inbox remains in use today.

Emergence of Autonomy and Scalability

As Omada continued to grow, care teams could reach hundreds of thousands of members. Omada’s datasets grew to represent an increasingly diverse member base, and coincidentally, statistical machine learning capabilities evolved. Thanks to more data at their disposal, coaches benefited from better data-driven insights on members, including improved predictive analysis, trend forecasting, member segmentation, and scoring of interventions. 

“With increased emphasis on the space for coach autonomy came the challenge of how best to review and performance manage coaches,” Quan said. Omada dedicated resources to determining how to make health outcomes more visible to coaches and tying coach-member engagement more clearly to clinical outcomes. Or, as Quan put it: “You can't just give somebody freedom, and then not tell them where to go.” 

You can't just give somebody freedom, and then not tell them where to go.”

In time, a clear picture began to emerge that showcased the power of AI to support––not replace––human-led care teams through the process of forming personal relationships with members: 

The ability of Omada’s human-led care teams to work in concert with different forms of supportive AI set the stage for coaches to help improve the technology they once fought against. 

AI as an Efficiency and Diversity Ally 

The coaches’ embrace of AI created new possibilities, including making better use of their most precious resource––time. This, Quan and Ellsworth agreed, was a key element in the scalability equation, as the bandwidth of coaches is a finite resource. In too many cases, health coaches functioned as tech support, in addition to everyday coaching duties. Even sifting through content libraries to identify what’s most relevant to members can be cumbersome. 

In the first version of our data science program, we failed to emphasize for our coaches the space for what they’ve told us was most meaningful to them, which is their autonomy.”

“In the first version of our data science program, we failed to emphasize for our coaches the space for what they’ve told us was most meaningful to them, which is their autonomy,” Quan said. “We should seek to understand what the core aspect of the human experience is for that coach-to-member relationship. While we must preserve that, and the coach’s clinical discretion, we can amplify the human experience, and then try to automate more mundane or time-consuming aspects at the edges of that, which doesn’t feel like a meaningful loss to the coaches.” 

AI-supported tools helped perform tech-support-like work, such as reminding coaches about members’ birthdays, reminding members about educational resources, and suggesting the timing of interventions. This freed up the bandwidth of coaches and allowed for wider reach. With more time on their hands, what do health coaches do? Besides reaching more members, Omada honed in on amplifying the creative potential and diversity-of-thought of over 270 coaches, each with unique member experiences. 

The idea is that the coaches are the masters of their craft, and it’s important to Omada that they have direct input in the algorithms that make suggestions to support them.”

“The idea is that the coaches are the masters of their craft, and it’s important to Omada that they have direct input in the algorithms that make suggestions to support them,” Quan said. “As they see what’s working in their own experience, they should add to the playbook, help evaluate it at scale, and surface the knowledge that’s most useful for other coaches to share with members. They inform us what’s working, and we’re helping get the right information to the right coach at the right time.”

Fostering Creativity and Growth with AI

One of the examples of human-led care teams working symbiotically with AI is Omada’s Coach Plays feature. This collaborative approach allows coaches to share their experiences and successful interventions. By creating a “playbook” of interventions and formalizing them into a configurable intervention management system, care teams can leverage that playbook to design and deploy their own specific messages and approaches for different types of members more efficiently, which helps to scale. The result is a rich, growing knowledge base that benefits algorithmic models, coaches, and potentially the entire Omada member population.

“We strive to find the balance between the two––AI and creative health coaching,” Ellsworth said. “At the freeform coaching stage, it was very rewarding. As a coach, I could just practice my craft, but I couldn’t adequately reach all the people who needed my help.” 

At the freeform coaching stage, it was very rewarding. As a coach, I could just practice my craft, but I couldn’t adequately reach all the people who needed my help.”

With better AI and a growing repository of personalized approaches at their disposal, we believe Omada coaches are empowered to become even more efficient health professionals who help members build belief and aim to drive long-term behavior change. Our data tells that story well. A 2022 analysis showed that diet quality improved by 38% when coaches–-supported by Omada’s data science program–-proactively recommended a specific healthy substitution, which can be drawn from the aforementioned “playbook.”

Omada coaches, Ellsworth said, now cherish the opportunity to scale their work meaningfully. “We believe coaches take a job at Omada as a digital coach because they can use their skill set and reach more people,” he said. “This technology helps them level-up their expertise by understanding population health science and impacting the whole member base.”

Looking Ahead

Omada’s journey with data science and even AI is far from over. We plan to continue taking all of the valuable lessons we’ve learned over the last decade and the enormous data corpus we’ve collected to start leveraging Generative AI (GenAI) and other state-of-the-art machine learning techniques in support of our care teams. We aim to advance personalization and member experience in new ways for greater impact. 

Our hope is that GenAI and deep learning–i.e. image recognition and large language models––will bolster our ability to support our human-led care teams in predictive and generative ways. 

Based on our years of experience, we believe in harnessing the power of AI technology to support human-led care, not replace it. That’s how we plan to empower our care teams and specialists to do what they do best: seek to move the needle on health outcomes. 

Omada is transforming healthcare through data.