R&D Lab

Adaptive adherence: getting digital health interventions actually used.

An open research line at Vuola.tech. It is not a product, it is a hypothesis we are working to prove or disprove. We are looking for a clinical partner, real usage data, and consortia where this belongs as a work package.

The problem

An intervention that is abandoned is a treatment that was never delivered.

Digital health interventions pass their clinical trial and then fail in the real world. Not because the clinical content is wrong, but because patients stop using them long before receiving the therapeutic dose. Without sustained use there is no outcome; without outcome there is no evidence; and without evidence there is no reimbursement.

3.3%
Median 30-day retention across 93 mental health apps, measured on real usage data, not developer reports
18.7%
Median dropout in randomised trials of app-based interventions for chronic disease
5-87%
Range of attrition reported across those same trials, an unexplained spread of that size is itself the finding

Sources: Baumel et al., Journal of Medical Internet Research, 2019;21(9):e14567. Rates of Attrition and Dropout in App-Based Interventions for Chronic Disease, JMIR, 2020;22(9):e20283.

Why it is not solved

Two opposite causes produce the same number.

The standard response is volume: more reminders, more notifications, more gamification. And when the industry says personalisation, it usually means segmenting by diagnosis or demographics and then sending everyone the same thing with their name on it.

One patient is overwhelmed by information and leaves. Another is under-stimulated and leaves. In the retention table both appear as a single figure, so the aggregate metric hides the mechanism. And because the default answer is to add more, every attempt to improve retention actively worsens the first group.

Two cancer survivors with the same diagnosis and the same app: one wants to explore their full record, the other closes the app on finding fifteen documents about recurrence. That is not a clinical difference. It is a difference in information appetite, and today's products do not model it.

Our hypothesis

Part of the unexplained variance in abandonment is a dosing mismatch, not a motivation problem.

Specifically: that a significant share of abandonment corresponds to a mismatch between the dose of information and novelty a product delivers and what each individual patient tolerates and wants; that this appetite can be inferred passively from in-product behaviour; and that calibrating it automatically improves measured adherence in a clinically relevant way.

01

Passive inference

Signal taken from in-product telemetry alone: what is opened, how deep the user goes, what is skipped, return latency, point of abandonment. No additional questionnaires, since questionnaire length is itself a known driver of dropout. No external or social network data.

02

Calibration

Three dosing parameters per patient: how much content, how novel relative to what they already know, and how fast to progress.

03

Measurement

A validated instrument for perceived experience, not just interaction metrics. Proxy metrics for novelty and serendipity are known not to align reliably with real user perception, which is an open problem in the field and the part we are best equipped to address.

04

Explainability

Every calibration decision traceable and auditable, a requirement for AI systems that adapt health interventions under the EU AI Act.

Where this stands

TRL 2. Concept formulated, not yet proven.

We are stating that plainly because the next step is a retrospective study on usage data that already exists, designed to be cheap and decisive. A negative result is a useful result and would close the line without committing development. This is a research line, not a product on sale.

What it rests on, all of it work done before Vuola: a doctoral thesis at the Universitat Politècnica de València on calibrating exploration according to the user's psychological profile, 17 peer-reviewed publications and more than 100 citations, an evaluation protocol for perceived accuracy, novelty, diversity and serendipity validated with real users, and product delivery, by our co-founder, of a digital health prototype validated with 92 patients and 16 healthcare professionals across four countries in the EU4Health smartCARE consortium.

PhD, UPV · recommender systems17 peer-reviewed publicationsValidated evaluation protocolEU4Health delivery experiencePIC 862917918
What we are looking for

Three kinds of partner.

01

A clinical partner, or a digital health company with usage data

Someone with a real intervention and a real abandonment problem, willing to let us test the hypothesis retrospectively on telemetry that already exists. This is the fastest and cheapest way to find out whether the signal is there at all.

02

A research group

Particularly in human-centred computing, digital health or recommender systems. We are a small company with a defined research question and published background, not a laboratory. The complementarity is obvious in both directions.

03

Consortia where this is a work package

Any European project deploying a digital intervention with patients has the same gap: nobody owns making sure it keeps being used, and without that there is no evidence to report. We can bring this as background and take that work package.

Does this problem sound familiar?

If you run a digital health intervention and abandonment is costing you outcomes, or you are shaping a proposal where this belongs as a work package, we would like to talk. We are equally interested in being told the hypothesis is wrong.

Talk to us about it