‘Aha Moments’ in Digital Health
How to identify engagement metrics that impact clinical outcomes
Welcome to the first issue of Behavioral Health Product Management, a monthly newsletter designed to help teams build engaging, clinically effective, and commercially successful digital health interventions.
If you’re receiving this email, it means you’re someone who I’d like to keep in touch with and who I think might be interested in the subject matter. That said, I understand your inbox is probably already a busy place, so if you’d like to opt-out, feel free to use the unsubscribe button at the bottom of this email.
If you enjoy the newsletter, please share it with others. I’d also love to hear your thoughts about the article, and what other topics you’d like to learn more about!
‘Aha Moments’ in Digital Health: How to Identify Engagement Metrics that Impact Clinical Outcomes
Engagement is arguably the single most important challenge in digital health today. Yet, while we recognize that engagement is important, there is a distinct lack of clarity regarding what exactly engagement is, and even more importantly, how to measure it.
Increasingly, digital health technologists and researchers agree that we need to find measures of meaningful engagement. That is, we must avoid the temptation to immediately use generic measures like number of sessions, weekly active usage, or program completion, and instead follow a data-driven approach to find the specific engagement metrics that uniquely predict long-term value in each intervention.
The quest for meaningful engagement metrics is not unique to digital health. In fact, many of the most successful technology companies have sought out such a metric, which when found, is closely linked to their ‘aha moment’. For example, Facebook identified ‘number of friends’ as its key engagement metric, and found that long-term retention was greater when users added 7 friends in 10 days. Slack, Twitter, Zynga, LinkedIn, and Twitter have all shared similar aha moments. A key principle is that meaningful engagement metrics are usually specific to the product. While it’s certainly possible that a generic measure of engagement could also be meaningful, it is rarely the case.
Finding ‘aha moments’ can be decomposed into two parts. First, you must find meaningful measures of engagement (e.g. friends), and second, you must identify critical levels of those engagement metrics that, once achieved, predict long-term user value (e.g. 10 friends in 7 days). In this article, I’ll cover the first step, finding meaningful measures of engagement. I hope to cover the second step in a future article.
Drawing from work in the consumer and SaaS industries, I’ve identified a three-step process for finding meaningful engagement metrics in digital health interventions.
Identify an indicator of value. For most consumer and SaaS products, it’s retention. For most digital health products, it’s clinical outcomes.
Identify potential leading indicators of that value. This is an iterative, exploratory process of generating competing hypotheses.
Determine whether the indicators are causal or correlational. Finally, you must validate your hypotheses by confirming or rejecting causality.
Step 1: Find the indicator of value
Indicators of value are one area in which there is a fundamental difference between digital health interventions and most consumer or SaaS products. For example, in a consumer product like Facebook, users are looking to be entertained, and therefore retention (consistently coming back to the app) is an excellent proxy for that value. Similarly, in a SaaS product like Slack, users want to achieve a specific objective, such as communicating with their colleagues. Once again, retention (consistently coming back to communicate with colleagues) is an excellent proxy for that value.
However, for most digital health interventions, the primary indicator of value is clinical outcomes1. Users have health needs and utilize digital health interventions to address those needs. If users come back to the app every day for months (strong retention), but their clinical symptoms do not improve, then they haven’t received the intended value from the product.
Moreover, in most digital health products there are not one, but three key stakeholders — patients, providers, and payors — all of whom value improved clinical outcomes. In this regard, clinical outcomes are a unified indicator of value for all major stakeholders.
Let’s take the example of a digital health intervention for depression. Patients, providers, and payors all care about decreasing depressive symptoms. We’ll continue with this example below.
Step 2: Find leading indicators of value
Now that you’ve identified a measure of value, it’s time to find leading indicators of it. Once found, these will become your metrics of ‘meaningful engagement’.
At this stage, it’s important to resist the temptation to immediately use a generic engagement metric. As mentioned earlier, it’s certainly possible that a generic measure of engagement will be a leading indicator of value, but by and large, meaningful engagement metrics are unique to your product.
To find and validate leading indicators of value, I recommend following the three-step subprocess outlined below.
Step 2a: Identify initial candidates based on theory and external data
If you’re building a new product from scratch, then you’ll need to rely on theory or external data to generate your initial hypotheses. Many digital health interventions are digitized versions of face-to-face interventions, and you can look at the predictors of clinical outcomes in those face-to-face interventions.2
Let’s turn back to our example. Let’s say your digital intervention for depression is based on Cognitive Behavioral Therapy (CBT). You investigate what predicts positive clinical outcomes in face-to-face CBT, and identify multiple predictors, three of which are:
Showing up for weekly appointments
Doing assigned homework (e.g. completing thought records and cognitive reframing exercises)
Having a positive therapeutic alliance with the therapist.
Step 2b: Map the theoretical indicators to specific, measurable actions within your intervention
Continuing with our example, your next step is to explore digital analogs for each of the traditional predictors:
Showing up for weekly appointments corresponds to weekly active usage (a generic engagement measure).
Doing assigned homework corresponds to completing in-app exercises.
A digital analog of therapeutic alliance is harder to identify because it’s a psychological process and is therefore difficult to measure with behavioral metrics (e.g. passive logs). However, you could potentially include a questionnaire that measures therapeutic alliance as part of weekly symptom check-ins, or could even experiment with sporadic prompts within the app.
Step 2c: Validate the proposed relationships with internal data
Once you’ve developed your hypotheses, it’s time to look at the correlation between the hypothesized leading indicators and clinical outcomes. Let’s say you find that:
Weekly active usage has a weak, non-significant, relationship with clinical outcomes.
There is a moderate, statistically significant relationship between completing specific types of exercises within the app and clinical outcomes.
After exploring therapeutic alliance further, you realize that while it is theoretically promising, it’s too difficult to reliably measure. It’s therefore not feasible to use as a primary engagement metric.
Based on the above, you choose completing in-app exercises as the most promising engagement metric and move onto step 3.
Step 3: Determine whether the relationships are causal or correlational
At this point, you’re probably thinking “But wait, correlation doesn’t equal causation!”, and you’re absolutely right. That’s why in this next step you’ll determine whether the hypothesized leading indicator (e.g. completing in-app exercises) causes the desired outcome (improved depressive symptoms), or whether there is actually some third factor (e.g. users’ pre-existing motivation to change) that leads users to both complete the engagement metric and improve their clinical outcomes.
You can test for causality by shipping new product features that increase your target engagement metric (e.g exercise completion rate), and then see whether that leads to a corresponding improvement in your clinical outcome metric (e.g. PHQ-9). For example, you might implement additional reminder notifications for exercise completion, or design a reward system that incentivizes completion of exercises.
Once you ship the feature, you’ll first confirm that the engagement metric does in fact go up as a result of the feature. If the target engagement metric goes up, you will then check whether the correlation with clinical outcomes is maintained. If the correlation holds, then that’s great support for your choice of engagement metric. Congratulations, you’ve likely found a measure of meaningful engagement! 3
Putting it all together
Finding meaningful measures of engagement is just the beginning. There are several important next steps that I will touch on briefly.
Blending multiple metrics
There’s rarely only one metric of meaningful engagement in an intervention. Instead, there are probably multiple ones, in which case you’ll be better off combining them into a hybrid measure of engagement. For example, you might categorize someone as meaningfully engaged if they do any 2 out of a list of 5 leading indicators within the program in a given week. Or you might want to create a hybrid metric that blends various factors into a single weighted average engagement value.
Measures of meaningful engagement are also likely to differ across users. E.g., users with more severe symptoms might benefit from a different style of engagement than users with mild-to-moderate symptoms. Optimal engagement style is also likely to depend on users’ pre-existing motivation levels, and may even change for the same user as they progress through the intervention. SilverCloud and Microsoft recently published an article that outlined their use of machine learning to identify different engagement styles. One of the major advantages of digital health interventions over traditional therapies is the ability to collect large amounts of data, which in turn, enable the personalization of interventions to match individual engagement characteristics.
Minimum effective dose
As alluded to at the start of this article, finding meaningful engagement metrics is only half of the equation. You must also determine the minimum effective dose of the engagement metric that predicts long-term user value. E.g., if you determine that completion of in-app exercises leads to improved clinical outcomes, you’ll next want to determine how many exercises should be completed over what time period in order to produce the greatest likelihood of clinical improvement.
Letting your metrics guide your product development process
Finally, once you’ve determined your meaningful engagement metrics and minimum effective dose, you should leverage your engagement toolkit to drive the specific engagement behaviors that lead to clinical outcomes.
If you’ve made it this far in the article, then there’s a good chance you’re as passionate about driving engagement in digital health as I am! I genuinely believe that it’s one of the most critical challenges facing our industry today (apart from perhaps commercialization). While there is no single, ‘magic’ engagement metric, the process outlined in this article will position you well to identify the meaningful engagement metrics that are an important milestone along the journey to building truly impactful interventions.
Three things I’ve read recently that shaped my thinking about digital health interventions.
This book is a classic of systems thinking that had been recommended to me by multiple people, and I’m glad I finally read it! From the internal systems of running a successful organization (research, product, design, engineering, sales), to the system that patients exist in (e.g. an adolescent living within their family, school, and social support systems), to the healthcare system itself (patients, payors and providers), systems are an integral part of almost every element of building a successful digital health intervention.
Reimbursement remains one of the most important challenges in digital therapeutics. With around 17% of U.S. adults and 35% of children covered by Medicaid (or CHIP), Medicaid reimbursement will be essential if digital therapeutics are to close the gap in health care access and equity. This article gave a great overview of how state Medicaid programs make coverage decisions and some of the challenges and opportunities for achieving reimbursement of digital health technologies.
Internet-Based Cognitive Behavioral Therapy for Depression
A Systematic Review and Individual Patient Data Network Meta-analysis
The question of how human guidance impacts clinical effectiveness in digital health interventions is an important one. As is common in social science, the findings tend to be conflicting, and this meta-analysis is the latest to tackle this issue. The analysis concluded that guided interventions were associated with greater clinical efficacy than unguided interventions, but that unguided interventions were still effective, particularly for those with less severe symptoms. Given that it is much easier and less costly to scale unguided interventions, the question should no longer whether or not interventions need to be guided in general. Rather the question is how to match the right style of intervention to the right type of patient to maximize impact.
Thanks to: Jim Liu, Mel Goetz, Elise Vierra, and Jessica Lake for their contributions to this post.
One notable caveat is for wellness products, like meditation apps, for which retention is arguably the best indicator of value. For example, if a user downloads a meditation app and comes back consistently over time, then the product is probably giving them their desired value. Also, since wellness apps generally have direct-to-consumer business models, you don’t need to be as concerned about payors’ and providers’ emphases on health outcomes. In such cases, you may be better off using retention rather than clinical outcomes as your primary indicator of value.
Leading indicators of clinical improvement are also closely related to the concept of active ingredients. Just as there are active ingredients in traditional pharmaceuticals surrounded by the remainder of the pill — which serves as a delivery mechanism — so too, there are elements of digital interventions that are critical for clinical improvement that are surrounded by the remainder of the software. In fact, there is a whole field of research into mechanisms of change within psychotherapies that seek to identify active ingredients, and such research can provide a great starting point for identifying what may be the leading indicator of clinical improvement within your digital intervention.
An important caveat here is that many digital health interventions do not have sufficient data to test for causality. For example, if your product is a prescription digital therapeutic going through initial clinical trials, then you’re unlikely to have a large enough sample size to draw meaningful conclusions about the relationship between individual engagement metrics and clinical outcomes. If that’s the case, then you’ll have to stop at step 2. However, if you’ve identified a moderate-to-strong correlation that is supported by a robust theoretical rationale for causality, then you can have some confidence that you’ve probably found an important relationship.