SmartCHANGE: Forecasting risk, changing behaviour, preventing Disease

SmartCHANGE is building something that does not yet exist in paediatric medicine: a robust, explainable AI risk prediction system that helps health professionals identify which children, based on their current behaviours such as diet and exercise, are most likely to develop chronic disease so they can intervene before it becomes inevitable.

“The problem is that children often behave in unhealthy ways and, of course, down the line, this leads to all sorts of chronic diseases,” says Mitja Luštrek, Head of the Department of Intelligent Systems at the Jožef Stefan Institute and project coordinator. “What we are trying to do is build a tool that allows doctors to evaluate better than they currently do whether a given child or adolescent is at risk and whether something should be done.”

He is not dramatising the issue here, but the implications are stark. The chronic diseases he is referring to are not rare or marginal conditions, but cardiovascular and metabolic diseases – “these are the conditions that kill about a third of people, probably even more,” he explains. Crucially, they are also largely preventable.

The logic behind SmartCHANGE is therefore deceptively simple: if cardiovascular and metabolic diseases are both widespread and preventable, then prevention must begin early, long before cholesterol levels, blood pressure and insulin resistance are firmly entrenched. Yet prevention in childhood has remained underdeveloped, particularly when it comes to systematic risk assessment.

For adults, predictive tools are abundant. “There are probably dozens of cardiovascular risk calculators, several for diabetes and many for other diseases,” Luštrek says. “But there is nothing comparable for children.”

This gap between what we can measure in adulthood and what we understand in childhood became the foundation for SmartCHANGE. The project’s central ambition is to create a clinically usable system that enables health professionals to evaluate risk in children and adolescents more accurately and more consistently than they currently can.

He is quick to point out that this is not simply about raising awareness. Many clinicians and parents already know that physical activity, diet and sleep matter when it comes to children’s health. The challenge lies in the fact that recognising these behaviours as important is not the same as knowing whether a child’s lifestyle actually meets recommended standards.

“Research has shown that many health professionals are unfamiliar with guidelines,” Luštrek explains. “Most of them do not know the guidelines for physical activity, for example. Parents may also misjudge their child’s health and many think their own children are healthy and leading healthy lifestyles when, in reality, they are not.

“They cannot judge whether a child’s behaviour is appropriate or whether it requires change. Deciding what that change should be comes later.”

For SmartCHANGE, this gap in perception is exactly where prediction becomes critical. The project is not simply encouraging healthier lifestyles in general terms; it is building a structured way to connect specific childhood behaviours with measurable future health risks.

The model is grounded in behavioural data because these everyday habits – physical activity, diet and sleep – are what ultimately shape long-term cardiovascular and metabolic outcomes.

By embedding that connection between daily behaviour and future disease directly into clinical practice, and combining systematically collected behavioural data with explainable analytics, SmartCHANGE aims to forecast risk in a way that is both scientifically rigorous and practically usable.

At the behavioural level, the focus is deliberately conventional. “Physical activity, diet and sleep are the three main behaviours we have included,” Luštrek says. “We also touch on screen use, and we include some mindfulness in our application, but these are not the core elements.

“Physical activity, diet and sleep are not novel risk factors either; they are the foundational determinants of long-term cardiovascular and metabolic health. Our aim is to identify risk through current behaviour and to change that risk by recommending changes in behaviour that we know lead to healthier outcomes in the future.”

To deliver this in practice, the project has developed two interconnected applications: one for children and families, and one for health professionals. The two are designed to work as a single system. The children’s app gathers behavioural data through short nutrition questionnaires, sleep tracking via wearable devices, physical activity measurement and self-reported fitness information.

“These data are combined into a continuous behavioural profile,” Luštrek explains. “When the child visits a health professional, further data are added: clinical measurements, health history, medications and lifestyle information. In some contexts, the system aims to connect directly to existing national fitness monitoring databases, streamlining data flow.”

In practical terms, a child’s daily activity and sleep patterns are tracked, nutrition habits are logged, and fitness levels are recorded. When they attend a consultation, whether with a paediatrician or a school nurse, the clinician accesses a dashboard that combines behavioural and clinical data into a single profile. From there, the system generates a structured risk assessment, highlights the behaviours driving that risk and allows concrete goals to be set, which are then transferred back to the child’s app for ongoing monitoring.

The methodological challenge at this point is clear: there are currently no long-term datasets tracking children from early life through to the eventual development of cardiovascular disease. “Because there are no proper risk models for children and no long-term datasets that track children all the way until they develop disease, we have to forecast risk factors.”

Rather than predicting disease directly, SmartCHANGE forecasts adult risk profiles. Using current child data, the system estimates what blood pressure, cholesterol, weight and related risk markers may look like at the age of 55. These forecasted values are then entered into established adult models, including SCORE2 and the Healthy Heart Score, along with a diabetes model.

The result is a quantified risk assessment, but one that does not operate as a black box, that is, a system that produces a score without revealing how it reached that conclusion. In many AI-driven systems, complex mathematical processes generate outputs that are difficult for clinicians to interpret or challenge.

A defining feature of SmartCHANGE, therefore, is its commitment to explainable AI. “We use explainable AI techniques to attribute risk to behavioural factors,” Luštrek says. “So we can say, for example, most of the risk is due to low physical activity and some due to diet, while sleep risks are fine.

“Instead of simply presenting a number, the system shows which behaviours are contributing most strongly to predicted future disease, making the output both interpretable and actionable for health professionals.”

The system also generates “counterfactuals” – alternative versions of the same child’s profile with lower predicted risk but otherwise similar characteristics. “This helps the health professional understand what is driving the risk and suggests interventions to move the child into a lower-risk state.”

This emphasis on transparency emerged partly from empirical findings within the project. In one study, the team deliberately inserted errors into AI outputs to test clinical responses. “Doctors tended to trust AI too much. We intentionally inserted mistakes into outputs, and they did not flag them.” The experience reinforced that explainability is not only about interpretability, but about encouraging appropriate scepticism.

SmartCHANGE has also demonstrated that its modelling approach performs “significantly better” than simple risk-factor flagging, such as high BMI alone, though, as Luštrek acknowledges, “it is still not perfect.”

While the health professional interface represents the core innovation, behaviour change ultimately depends on engagement, and engagement, particularly among children, requires a different strategy. “Everyone agrees prevention is important – except perhaps the kids!”

Participatory design workshops across four countries ensured children’s perspectives shaped development. When discussing risk communication, the consortium confronted a psychological reality: “Teenagers often believe they are immortal and discount negative outcomes, meaning fear-based messaging would be ineffective. Gamification then became central as kids don’t care about health, they care about fun.”

The result is the Happy Plant app. Children complete healthy challenges and receive “water” to grow a virtual plant; when fully grown, it joins a garden, and parents may link completion to real-world rewards. “Even adults who saw it said they would like to use it.”

Beneath the playful game lies a serious behavioural insight. Habits formed in childhood persist, and identity shapes action. “If someone identifies as a healthy person, they behave accordingly. We don’t only want to modify behaviours, but to reinforce healthy self-concepts during formative years.”

From both a technical and ethical perspective, SmartCHANGE has proceeded carefully. Accessing existing datasets required extensive data-sharing agreements, while all new project data needed formal ethical approval. Before any wider rollout, identifying information will be removed, and certain variables will be adjusted to reduce the risk of re-identification.

To further protect privacy, the project uses federated learning. Rather than pooling all data into a single central database, each country keeps its data locally and contributes to a shared model through privacy-preserving techniques such as differential privacy.

The system is also designed to evolve. Models can be retrained regularly and monitored for “concept drift”, ensuring predictions remain accurate as behaviours and population patterns change over time.

As the project moves through its feasibility study phase, now being tested in Taiwan as well as in the original four European countries, attention has turned to sustainability. Integration with national fitness monitoring systems and e-health infrastructures appears the most promising pathway. The risk prediction engine itself is “likely the main exploitable outcome.”

Yet perhaps the most grounded insight from SmartCHANGE is its realism about scale and pace. “Behaviour change is difficult and rarely dramatic,” says Luštrek. “Risk prediction alone will not transform public health overnight. But embedding it consistently within healthcare systems, small improvements may accumulate.

“Children today probably lead less healthy lifestyles, partly due to digital devices and reduced outdoor play. If habits persist into adulthood, and if identity solidifies early, then prevention must begin long before disease manifests,” he concludes.

SmartCHANGE does not promise immediate transformation. What it offers instead is a structured, clinically grounded system that connects childhood behaviour to lifelong health, making prevention actionable at the stage when it can still reshape a person’s future rather than manage its consequences.