The development of risk-prediction models started with the collection of data. We obtained 17 datasets, most of which have data of several thousand children containing diverse health-related variables, from biological such as blood pressure, to behavioural such as physical activity. We harmonised these datasets and we augmented them with synthetically generated data: when a dataset did not contain a key variable, we generated the missing values by following the patterns in other datasets where this variable was present. This enabled us to build a model than can forecast all relevant risk factors from childhood to adulthood. This way we can paint a picture a child when they become an adult, and we can then use existing risk models designed for adults to predict their NCD risk.
We also worked on making our AI solution more trustworthy. We extended the models with the ability to express confidence in predictions, which empowers health professionals to make informed decisions. We developed a counterfactual generator: a tool that can generate hypothetical scenarios to show how different risk factors and behaviours can change the predicted health risk. Such a tool can help users understand the model's decisions to build trust, and suggest ways to lower risk. We made it work in a federated setting, meaning that not all data is at one site, but rather that each sites stores its own data to preserve privacy. While we believe the counterfactuals are useful in SmartCHANGE, they can be difficult to understand, so we developed a tool to visualise them in a way useful for health professionals.
Most of the participatory design activities have been completed. Design sessions with adolescents, family members and health professionals were organised to better understand their needs and preferences. Different design directions were explored and clickable prototypes validated. Key insights include the families’ and adolescents’ desire for a holistic application that addresses multiple health areas expressed, and health professionals’ need to be guided in helping families and adolescents towards lifestyle changes. Since the SmartCHANGE solution is intended achieve healthier behaviour, which is notoriously difficult, we investigated the existing literature on what interventions are most effective. We identified five key principles: enhance inner motivation by supporting personal reasons and values for healthy behaviour, support feelings of competence by giving positive feedback, promote autonomy and opportunities to choose own actions, enhance positive emotions, and encourage relatedness to others.
Building on the results of the participatory design, we are developing applications for health professionals and citizens. The application for health professionals is focused on displaying the risk and the counterfactuals. The application for families, HappyPlant, encourages healthy lifestyle changes using a plant growth analogy. Personalised goals, focusing on movement, nutrition, sleep, and mindfulness, are set by the SmartCHANGE models and adjusted by health professionals. Users have the autonomy to select goals to commit to. Subsequently, when goals are met, the plant thrives, incorporating gamification elements. Additional components include a balanced multi-layered reward system focussing on short- and long-term goals. This approach, which focuses on the strengths of the children, received most favorable feedback from both target groups.
Since health data, especially of children, is highly sensitive, and AI is currently in the focus of regulators, we have a legal expert in the project who continuously monitors our work to make sure we adhere to all pertinent laws and ethical standards. Once our applications are complete, we will conduct a small pilot study mimicking real usage as much as possible. We have already prepared the protocol for this and are currently waiting for approval of ethics committees at the pilot sites.
