When pediatric pulmonologist Elin Kersten joined the first cohort of the exquAIro Biomedical AI Bootcamp, she did not yet have a fully formed research line in artificial intelligence. What she did have was a clinical question that kept returning in her daily practice at the University Medical Center Groningen (UMCG): which children with early respiratory symptoms will go on to develop asthma and which will not? “This project actually started during the bootcamp,” she says. “That’s where I first put it on paper.”
Earlier this year, that idea translated into a Longfonds grant (the Rutgers Kickstart Subsidie), enabling her to develop an independent research line focused on the earliest origins of childhood asthma. The project combines multi-omics data, environmental exposures and machine learning to identify predictive signals of disease, long before diagnosis becomes possible.
A clinical question, reframed through AI
Kersten’s work sits at the intersection of patient care and data-driven research. As a pediatric pulmonologist, she primarily sees children with severe or complex lung conditions referred to the UMCG. Alongside this, her research focuses on the early development of asthma, particularly in very young children presenting with recurrent respiratory symptoms. “In the UMCG, we see these children early on,” she explains. “But at that stage, it’s often unclear who will actually develop asthma later.”
The bootcamp played a key role in reframing that clinical uncertainty into a structured research problem. Rather than approaching AI as a purely technical tool, the program emphasizes identifying opportunities, defining the right questions, and designing projects that can ultimately have impact.
The nose as a data-rich interface
The project is embedded in Lifelines NEXT, a large longitudinal birth cohort in the Netherlands. In over 1,300 infants, nasal brush samples were collected at 3 and 12 months of age, alongside detailed environmental data gathered during pregnancy and early life.
From these samples, two key biological layers are derived: the composition of the nasal microbiome and the gene expression activity of the nasal mucosa. This allows Kersten to study not only which microorganisms are present, but how the host tissue responds. “We’re measuring RNA,” she explains, “so you’re looking at which biological processes are active in the cells at that moment, not just genetic predisposition.”
These early-life data are then linked to outcomes at school age, where asthma is assessed through clinical measures including spirometry. The aim is to identify combined microbial–host signatures that precede disease development.
When scale and structure require new methods
What makes this problem particularly suited to AI is not just the volume of data, but its complexity. Each child contributes thousands of variables across multiple domains, resulting in a dataset of roughly a million data points. “All those variables are connected,” Kersten says. “They’re not independent, so you can’t just analyse them with classical statistical models.”
Instead, the project explores a range of machine learning approaches, comparing different model types and iteratively refining them. The process is as much about model selection and validation as it is about prediction itself. Care is taken to avoid overfitting, using separate training and validation datasets and, ultimately, testing performance on an external cohort from the United Kingdom.
From prediction to understanding
While the development of predictive models is a central component, the broader aim is to better understand the biological mechanisms underlying asthma. The study connects to the biodiversity hypothesis, which suggests that reduced microbial exposure in early life may increase susceptibility to allergic disease.
By linking microbial composition to gene expression in the airway, Kersten hopes to move beyond general associations toward more precise insights. “If we can identify early markers,” she says, “you could potentially recognise at-risk children much earlier and maybe even intervene.”
Learning to bridge disciplines
For Kersten, one of the most important outcomes of the exquAIro bootcamp was not just technical familiarity with machine learning, but the ability to engage across disciplines. “It wasn’t part of my medical training,” she says. “But you learn that those complex models are built on concepts you already understand.”
This has changed how she collaborates with data scientists and analysts. She is now able to ask more targeted questions, critically assess modelling choices, and contribute as an equal partner in interdisciplinary teams. This has become an increasingly important skill in modern biomedical research.
A more critical use of AI
At the same time, that experience has made her more selective in how AI is applied. She points to a growing number of studies where machine learning is used without a clear path to implementation.
“It’s nice to develop a model,” she says, “but the question is what you’re going to do with it afterwards.” For Kersten, the value of AI lies in its ability to uncover meaningful patterns that can inform clinical decisions, not in the model itself.
Toward earlier intervention
By combining clinical insight, multi-omics data and machine learning, this project represents a shift toward earlier detection and, potentially, prevention of asthma. Instead of diagnosing disease after symptoms appear, the aim of Kersten’s research is to identify risk profiles in infancy, when interventions may still be possible.
More broadly, it illustrates how targeted AI training can accelerate that shift: not by replacing clinical expertise, but by equipping clinicians to translate complex data into new practices that benefit patients.