How to Test Healthcare Messaging with Synthetic Patients Grounded in Behavioral Science

A patient engagement team has three versions of an outreach message about bowel cancer screening, and the send date is Friday. Testing with real patients would take weeks, and the list is too large to learn from a bad send. The team needs a first filter that removes the weakest versions before anyone is contacted.

Synthetic patients are model-generated personas that react to a message the way a defined type of patient might. Grounding them in behavioral science means each persona carries explicit behavioral variables, such as what the person knows, what their circumstances allow, and what motivates them, instead of a demographic label alone. This article shows how to build such a panel, run a message through it, and decide what still needs real patients.

Why does healthcare messaging need behavioral science?

Healthcare messages succeed or fail on behavior, and behavior depends on more than age and zip code. Two patients of the same age can read the same reminder and act differently because one understands it and the other does not.

The evidence that wording matters is clinical. A systematic review of 96 studies found that low health literacy was consistently associated with more hospitalizations, greater use of emergency care, less screening mammography, less influenza vaccination, and a reduced ability to take medications appropriately and interpret health messages correctly (Berkman et al., 2011). Messaging also moves attendance directly. A Cochrane review of 8 randomized trials with 6,615 participants found that, across the 7 trials comparing text message reminders with no reminder (5,841 participants), attendance rose from 67.8% to 78.6%, a risk ratio of 1.14, with moderate-quality evidence (Gurol-Urganci et al., 2013).

Those results explain why teams test copy before sending it. Wording changes what patients do, so a weak version costs real outcomes.

What is COM-B, and how does it shape a synthetic patient?

COM-B is a model of behavior with three parts: capability, opportunity, and motivation. Michie, van Stralen, and West (2011) built it while reviewing 19 existing behavior change frameworks. Capability is a person’s mental and physical ability to do the behavior, including knowledge and skills. Opportunity covers the factors outside the person that make the behavior possible or prompt it. Motivation covers the brain processes that energize and direct behavior, including habits, emotions, and conscious planning.

For persona design, COM-B turns a vague profile into a set of testable conditions. Compare two cards for the same screening message:

The same text can succeed with Persona B and fail with Persona A. A panel that varies these three fields exposes that gap before a send.

How do you test healthcare messaging with synthetic patients?

The method has five steps, and the first two decide most of the quality.

1. Define the target behavior. Write one sentence naming who should do what by when, such as “Adults aged 50 to 74 book a screening within 30 days.” A message with two behaviors gets tested against neither.
2. Build a panel that varies COM-B, not only demographics. Include personas with limited health literacy, low opportunity, and low motivation, since they are the ones messaging usually fails. Platforms for AI user research for healthcare can generate such panels from a target profile in minutes.
3. Show each version and ask open questions. Ask what the message is asking them to do, what would stop them, and how it makes them feel. Avoid asking which version they like, because simulated respondents agree too easily.
4. Code every barrier to COM-B. Label each objection as a capability, opportunity, or motivation problem. A capability barrier calls for simpler wording, an opportunity barrier for a different channel or time, and a motivation barrier for a different framing.
5. Revise and retest. Fix the top barrier, run the new version through the same panel, and compare. Stop when the same objections stop recurring.

What does a coded result look like?

A coded result maps each barrier to a specific edit, so the next draft changes for a reason. The examples below are hypothetical.

Keep the coding table with the final copy. Reviewers can then see why each edit was made.

What are the limits of synthetic patients?

Synthetic patients are a filter. They do not replace patients, and several kinds of evidence say so. In a study spanning 3,200 human participants, 16 demographic identities, and 4 language models, Wang et al. (2025) found that models tend to misportray demographic groups and flatten the variety within them. The authors urge caution wherever identity matters to the task, which describes most health communication.

Nielsen Norman Group reached a similar position for synthetic users generally (Rosala & Moran, 2024). It found that they gave overly optimistic accounts of their own behavior and “seem to care about everything,” so priorities come out flat. It also stated that presenting synthetic findings as real-user findings is unethical.

Three safeguards follow. Use synthetic results to choose between drafts, never to approve safety-critical content. Disclose which findings are simulated in any internal or published report. Run the finalist with real patients and community reviewers, especially for groups the model may flatten.

How do you confirm the winning message with real patients?

Confirm the finalist with a small round of real patients, then measure it live. A small group of patients from each key segment, interviewed about the finalist alone, can surface problems a simulation missed. After launch, compare the new message with the old one on the behavior you defined in step 1, using a sample size calculated in advance for the lift you expect. Record the gap between what the panel predicted and what happened. That gap tells you how much weight to give the next simulated result.

References

1. Berkman, N. D., Sheridan, S. L., Donahue, K. E., Halpern, D. J., & Crotty, K. (2011). Low health literacy and health outcomes: An updated systematic review. Annals of Internal Medicine, 155(2), 97-107.
2. Gurol-Urganci, I., de Jongh, T., Vodopivec-Jamsek, V., Atun, R., & Car, J. (2013). Mobile phone messaging reminders for attendance at healthcare appointments. Cochrane Database of Systematic Reviews, 2013(12), Article CD007458. https://doi.org/10.1002/14651858.CD007458.pub3
3. Michie, S., van Stralen, M. M., & West, R. (2011). The behaviour change wheel: A new method for characterising and designing behaviour change interventions. Implementation Science, 6, Article 42. https://doi.org/10.1186/1748-5908-6-42
4. Rosala, M., & Moran, K. (2024, June 21). Synthetic users: If, when, and how to use AI-generated “research.” Nielsen Norman Group.
5. Wang, A., Morgenstern, J., & Dickerson, J. P. (2025). Large language models that replace human participants can harmfully misportray and flatten identity groups. Nature Machine Intelligence. (Accepted; preprint arXiv:2402.01908)