Behavioral Design for Digital Therapeutics and Wellness Platforms

Last updated October 19, 2026

Independent usage data on 93 popular mental health apps, analyzed by Amit Baumel and colleagues in the Journal of Medical Internet Research study of real-world mental health app engagement, found that the median app kept about 1 in 30 of the people who installed it. That 30-day retention rate, the share of people still opening the app a month after installing it, was 3.3%, against 3.9% at 15 days.

The Decision Lab is a behavioral design firm that helps digital therapeutics and wellness platforms raise patient adherence and improve clinical outcomes while keeping patients safe. Through our behavioral product design practice, we pair research with patients and clinicians with product design that is tested before it is built. Our work covers chronic disease management, mental health, youth support, prevention and digital wellness. In that work, adherence improves when the product gets each patient to a first therapeutic action quickly, fits the care plan to what that patient can manage, sends few reminders tied to specific actions, judges success by health outcomes rather than time in the app, and hands patients to human care when the app cannot safely help.

Digital therapeutics are software products that deliver a treatment, such as a structured therapy program or a disease-management plan. For these products, improving adherence without over-notifying patients is the central design problem, since more prompts can raise app opens while leaving patients guilty about missed days or worn out by alerts, and neither outcome helps the condition the product exists to treat.

These published projects show the range of our behavioral design work in health. Outcomes are listed where the case study discloses them.

Product and clinical teams engage us for some or all of these services:

1. Patient and clinician research. Interviews and surveys with the people who use the product and the clinicians who prescribe or support it.
2. Behavioral diagnosis and barrier framework. A map of what stops patients from starting and continuing treatment, with each barrier matched to behavior change techniques that have evidence behind them.
3. Onboarding and care-plan adherence design. First-session flows built backward from the first therapeutic action, plus responses matched to the reason a patient missed a dose or step.
4. Notification and intervention strategy. A reminder budget with rules for timing and framing, including when to send fewer prompts.
5. Safety exclusions and escalation pathways. Clinician-defined exclusions for personalization, plus escalation triggers and routes to human support.
6. Experiment design and product validation. Tests of target behaviors before production code is written, with controlled experiments where they are feasible.
7. Measurement plans. Metrics that link engagement data to the clinical or behavioral outcomes the product is meant to change.

Most engagements start with the first two services and move into design once the barriers are clear. We test design decisions before development wherever we can. For a global consumer electronics company, every target behavior in a new mindfulness service was tested through experimentation before production code was written, and we coached the client's team until they could run the process without us.

Controlled experiments are the strongest evidence for a design change, though they are not always possible or necessary at every stage. Behavioral design also does not replace clinical validation, and a product that makes therapeutic claims still needs the evidence its regulator requires.

Firms in this space split into those that design the behavior-change mechanism and those that build the whole platform, and many products need both. Whichever kind you hire, we would expect any firm on a shortlist, including us, to meet these five tests:

1. It researches your actual patients before designing, rather than relying only on published personas or general engagement patterns.
2. It works with your clinicians on exclusions and escalation rules, and treats patient safety as a design input.
3. It tests target behaviors before production code is written, with controlled experiments where they are feasible.
4. It defines success in clinical or behavioral outcomes, with engagement metrics as supporting evidence.
5. It can show health products it has taken to launch, including what did not work along the way.

We design the behavioral layer of health products and have stayed through launch on projects such as PocketWell and the Kids Help Phone redesign.

1. Get the patient to a first therapeutic action in the first session, and move everything else later.
2. Find out why each patient misses doses or care-plan steps before choosing how to respond, since forgetting is only one cause.
3. Personalize by the patient's capability and readiness to change, then check every recommendation against their health context before it reaches them.
4. Treat reminders as a limited budget, and spend each one on a specific action the patient can take at that moment.
5. Track clinical and behavioral outcomes alongside engagement, and stop optimizing any engagement number that does not move them.
6. Define with clinicians when the app hands a patient to human care, and make that route visible at every point in the product.

The first therapeutic action is the earliest step in the product that has clinical value on its own, such as logging a first dose or completing a first therapy exercise. We design onboarding backward from that step. Anything that does not help the patient reach it in the first session, including optional profile fields and feature tours, moves to later sessions.

The Baumel figures show why the first session carries so much weight. Retention at day 15 and day 30 was almost the same, so in that sample nearly all of the loss had already happened within the first two weeks.

Goal setting during onboarding needs as much care as the flow itself. In our work on the MyHeart Counts Canada behavioral product design project, user research showed that guilt often derailed users' fitness goals and led them to avoid the app and then uninstall it. Unrealistic goals set at onboarding were an early source of that guilt. We recommended asking users to report their current activity level just before they set a goal, so the goal anchored near their existing habits while still asking them to stretch.

For a digital therapeutic, the same logic applies to dosing schedules and exercise targets. A first week the patient can complete gives the product evidence of progress to build on, while a first week they fail gives them a reason to stop opening it.

Most adherence features assume the patient forgot. The REMIND randomized trial of medication reminder devices, led by Niteesh Choudhry and published in JAMA Internal Medicine, tested that assumption with 53,480 patients who had a record of not taking their chronic disease medications. Patients were mailed one of several low-cost devices, from a standard pillbox to a bottle cap with a timer showing time since the last dose, and a control group received nothing. After 12 months, about 1 in 6 patients in each group had their medication on hand at least 80% of the time, and none of the devices raised that share over the control group.

We read that result as a diagnosis problem. Patients skip doses because of side effects, cost, doubts that the drug is doing anything, or a condition that causes no symptoms, and a reminder answers none of those. In our work with Chronwell on Livvy, a digital therapeutic for fatty liver disease, the core challenge was that patients find it hard to take a diagnosis seriously before they feel any symptoms. We surveyed about 800 people who had the condition or were at high risk of it, and used their answers to build a behavioral framework of the barriers that kept them from making lifestyle changes. Each barrier was matched to behavior change techniques with evidence behind them, and our design recommendations were low-cost changes to an already well-built app, aimed at those specific barriers.

Care-plan adherence covers more than pills, since booking follow-up appointments and doing prescribed exercises count too. The same diagnostic step applies to each. When a patient misses a step, the product should find out why, with a single low-effort question where possible, and route the response to that reason. A patient who missed a dose because of nausea needs a prompt to talk to their care team, while a patient who missed it while traveling needs help planning around the trip.

Personalization in consumer apps usually means changing content by interest. In a digital therapeutic, the variables that matter are what the patient is able to do and how ready they are to change, and the patient's health context sets limits on both.

Capability covers health literacy, eyesight, dexterity and the amount of treatment a patient is managing at once. When the HeartStrong trial found no effect from connected pill bottles and daily lotteries in heart attack survivors, the authors raised the possibility that taking several medications, or using the electronic bottles themselves, made adherence feel daunting. A plan that adds steps for a patient already juggling several prescriptions can lower adherence even when every step is well designed.

Readiness changes over time. A newly diagnosed patient may still be deciding whether the diagnosis is serious, while a patient several years into treatment may be tired of it. In our work on the Kids Help Phone youth mental health platform, the audience ran from elementary school children to young adults, including newcomers to Canada and Indigenous youth, and our research with hundreds of young people across the country shaped how the platform offered information at different levels of support. Our UC Berkeley digital mental health work started from the same premise, that the support people need varies with their background and changes as their mental health changes.

Health context is where personalization becomes a safety question. A goal or reminder that suits most users can harm a patient with a particular history, such as weight-loss targets for someone with a history of disordered eating. We recommend that clinicians define those exclusions before launch, and that the personalization logic checks them before anything is shown.

Reminders work best as a limited budget. In the HeartSteps micro-randomized trial of mobile activity suggestions, Pedja Klasnja and colleagues sent 44 adults up to five suggestions a day for six weeks, randomizing at each point whether a suggestion went out. Early in the study, a suggestion raised the number of steps taken in the following 30 minutes by about two thirds compared with no suggestion. The effect shrank as the weeks went on, and averaged across the whole study it fell to 14%, too small to rule out chance.

How a reminder is framed matters as much as how often it arrives. A megastudy of text message nudges for flu vaccination, led by Katherine Milkman in the Proceedings of the National Academy of Sciences, tested 19 messages with 47,306 patients before primary care visits at two health systems. On average the texts raised vaccination rates by about 5% over patients who received no study text. Messages did best when they told patients a vaccine was reserved for them and sounded like the kind of message patients expect from their doctor's office, while casual or interactive messages did worse.

From this evidence and our own design work, we apply a few rules to reminders in digital therapeutics:

• Every reminder names one action the patient can take right now, such as taking the evening dose.
• Reminders follow the patient's chosen times, and the patient can pause or snooze them without leaving the screen.
• Ignored reminders reduce future volume. A run of dismissed alerts is a signal to send fewer, or to ask what is getting in the way.
• Tone matches a care provider. No streak counters that reset to zero, and no messages that frame a missed dose as a personal failure, since guilt was the pattern that drove MyHeart Counts users to uninstall.

Incentives need the same caution. In the HeartStrong trial of reminders and incentives after heart attack, led by Kevin Volpp, 1,509 heart attack survivors were randomized to usual care or to a program built around connected pill bottles and daily lotteries. Each adherent day earned a lottery draw with roughly a 1 in 5 chance of $5 and a 1 in 100 chance of $50. Average adherence was 46% in the program group against 42% under usual care, a difference too small to rule out chance, and readmissions for heart problems did not differ. When a reward is tied to a logged action rather than a health outcome, we suspect it also invites patients to log doses they did not take, which corrupts the data the care team relies on.

Engagement metrics tell a team whether patients are using the product, and clinical metrics tell them whether the product is helping. In a consensus paper on effective engagement with digital behavior change interventions, Lucy Yardley and colleagues argued that the target should be effective engagement, defined as enough engagement to achieve the intended outcome, rather than as much engagement as possible. A patient who completes a 10-minute exercise twice a week and gets better is a success, even if a dashboard shows them as a low-activity user.

The HeartStrong trial shows the gap from the other side. Its primary outcome was time to readmission for a heart problem or death, and the program changed neither that outcome nor adherence enough to detect. Activity inside the program, such as bottle openings and lottery draws, said little about the outcome the trial existed to change.

We pair each engagement metric with the clinical or behavioral measure it is supposed to predict, and we check that link in the product's own data before optimizing for it.

Every digital therapeutic needs a defined point where it stops trying to keep the patient in the app and hands them to a person. Common triggers include signs of crisis or self-harm, symptoms getting worse despite adherence, possible side effects, and requests the product was never designed to handle.

The cost of missing that point is visible in the case of Tessa, a chatbot the US National Eating Disorders Association had promoted as a prevention resource. In May 2023, the association took Tessa down after it gave weight-loss advice to an activist with a history of an eating disorder, including suggestions to count calories and aim for a daily calorie deficit. About 2,500 people had used the chatbot by then.

Several of our digital mental health projects have centered on making the route to human help easy to find at the moment it is needed:

• For PocketWell, the companion app to the Wellness Together Canada platform, we led research and product design, and the app gives mobile users quicker access to phone counseling and crisis support alongside its mood tracker.
• For Kids Help Phone, we led a redesign built on stepped care, in which young people can use self-guided resources or reach professional counselors and trained volunteer crisis responders, depending on what they need.
• For UC Berkeley, the brief included making it easier for students who need more intensive treatment to move from digital support to face-to-face services.
• For the Indiana National Guard, we studied the barriers that kept at-risk service members from seeking psychological treatment, including a culture of silence and fear of what disclosure would do to their careers.

In practice, we recommend that clinicians write the escalation rules before launch and that the route to human support stays one tap away on every screen. Once a patient has been escalated, the app should pause its engagement prompts. A patient in crisis should never receive a streak reminder.

What does a behavioral design firm do for digital therapeutics?

A behavioral design firm applies research on how people make health decisions to the product itself, from the first-session flow to reminder rules and escalation paths. At The Decision Lab, that usually means research with patients and clinicians followed by design changes tested before development, judged against clinical outcomes rather than app usage.

How many push notifications should a health app send?

There is no universal number, and the evidence favors fewer, better-timed reminders. In the HeartSteps trial, the effect of activity suggestions faded over six weeks. We set a reminder budget for each patient, tie every reminder to one action, lower volume when alerts are ignored and let patients pause reminders without leaving the screen.

Do financial incentives improve medication adherence?

Incentives sometimes help, but not reliably. In the HeartStrong trial, daily lotteries combined with connected pill bottles did not measurably improve adherence or readmissions among 1,509 heart attack survivors. Rewards tied to logged actions can also reward logging over taking medication, so we test incentives against clinical measures and prefer removing the barrier behind a missed dose.

What is the difference between engagement metrics and clinical outcomes?

Engagement metrics, such as daily active users or notification open rates, describe how patients use a product. Clinical outcomes, such as symptom scores or hospital readmissions, describe whether their health improves. The two can move apart, so we pair each engagement metric with the outcome it should predict and confirm that link in the product's own data.

When should a digital health app refer a patient to human care?

An app should hand a patient to a person when it detects signs of crisis or self-harm, worsening symptoms, possible side effects, or a request outside its clinical scope. Clinicians should define those triggers before launch. The route to human support should stay one tap away on every screen, and engagement prompts should pause after escalation.

If you are building or improving a digital therapeutic or wellness platform, talk with our behavioral product design team about patient research, adherence design, notification strategy and safe escalation pathways.