Leveraging GenAI to Optimize Behavior Change Predictions

By Sarah Linke, PhD, MPH

By now the evidence is clear that virtual health programs offer improved efficiency and convenience for managing chronic conditions. Less frequently discussed but equally important is that they also add a new dimension to measuring and analyzing patient behavior.

While traditional models of conceptualizing behavior change progression, such as through the Stages of Change, still apply, virtual settings offer billions of data points that can be utilized by machine learning (ML) and generative artificial intelligence (GenAI) to make predictions that inform personalized interventions.

Anticipating when patients are likely to experience setbacks or relapse into undesired behavioral patterns based on historical trends or similar patients’ data enables care teams to provide just-in-time interventions. For example, a well-timed message from a health coach checking in on goals, barriers, and action plans may help patients successfully navigate high-risk situations, which could be as common as overeating at weekend brunch.

Currently, ML technology informs behavior change predictions based on quantitative and chronological data; however, GenAI has the potential to take us even further. But what types of data are valid, pragmatic, and clinically meaningful in this context? And how can we use ML and GenAI to optimize behavior change predictions?

Key Data Inform Meaningful Predictions in Behavior Change

Ultimately, we know that the key lifestyle behaviors associated with better health boil down to the basics: healthy eating, physical activity, sleep, and stress management. We can measure these behaviors subjectively through patient self-reporting, and more objectively through wearable tracking tools, and with FDA-approved digital devices like those used in Omada’s programs.

However, there’s so much more data to pull from in a virtual setting. Sources such as backend platform data, wearables, electronic health records, surveys, and virtual interaction logs also provide treasure troves of data. They reveal patient insights, biometrics, adherence patterns, and sentiments, forming the foundation for understanding behavior change. A healthcare provider like Omada can use this data, as data in an EHR would be used, to better understand their patient, all while maintaining the patient’s privacy.

With this wealth of information at our fingertips, which data points should be used to make the most meaningful and scalable predictions? To start, specific types of health behavior engagement metrics come into play. These metrics may include (but are not limited to):

• Frequency of virtual interactions
• Biometric data
• Text data
• Engagement with specific platform features
• Self-monitoring

However, not all engagement metrics are created equally. Active forms of engagement–like setting goals, logging meals, and interacting with care teams and peers–tend to be more predictive of meaningful behavior change than passive forms of engagement–such as logging in or reading a lesson. Seeking social support, for example, has been associated with astounding improvements in health behavior change.

Analyzing and Interpreting the Data to Evaluate, Predict, and Intervene

With billions of health-related data points at our disposal, how do we use technology to extract meaningful insights that unveil behavior change? We have options depending on what we’re trying to achieve.

In the context of behavior change, it might look like the following:

1. ML categorizes patients into groups based on similar behavior trajectories.

2. Natural language processing (NLP), a branch of AI, parses patient communication to detect changes in sentiment, including subtle shifts in attitudes that may predict behavior changes.

3. GenAI distills the above data into a comprehensive text summary delivered to care teams, enabling them to efficiently and accurately make targeted interventions.

What could this look like in practice?

The holidays are coming up, and a provider may be interested in connecting her patient’s eating and activity patterns with her CGM data over time. She might also want to know her patient’s historical sentiment around the holidays. The provider prompts GenAI technology to condense these relative data points and form a personalized care plan to support her patient’s physical and emotional health among myriad holiday triggers, e.g. sugary foods, travel, and family stressors.

The provider can then send her patient tailored outreach messages about her behavior, such as “did you know that you tend to get fewer steps in November? What might be different about that month for you?” This type of predictive outreach, along with support, tools, resources, or reminders will help reduce the patient’s likelihood of falling back into patterns they are trying to avoid.

It also provides the patient with an opportunity for self-reflection and partnership with their care team to identify potential barriers and create an action plan to achieve their goals.

Integrating GenAI With Intention

How do we achieve this scenario in virtual care?

Successfully harnessing GenAI to predict and impact behavior change means actively imbuing it with clinical understanding. As we look toward the future, and iterate on our extraordinary and comprehensive data set, improving patient outcomes remains top of mind.

At Omada, we’ve been meeting our members where they are since the beginning. Advances in technology will help us stay a few more steps ahead, so we can usher them to a healthier destination.

This Proof Points edition was originally published on LinkedIn on 9/25/23.