AI Model Flags Which Food-Allergic Babies Will Develop Asthma by School Age

For parents of a baby newly diagnosed with a food allergy, one of the most anxious unanswered questions is whether the child will go on to develop asthma a few years later. Food allergy in early life is one of the strongest known predictors of childhood asthma, yet clinicians have lacked a reliable, non-invasive way to identify, during infancy, which of these children are truly on the path to the disease. A new prospective cohort study published in Pediatric Research by a team at Wenzhou Medical University and collaborating institutions now reports a machine-learning model that attempts to close that gap, screening infants between birth and age two and predicting their asthma risk at school age using routine clinical data alone.

The study enrolled 447 infants with confirmed food allergy, all aged between zero and two years, and collected an unusually rich baseline dataset: 135 candidate predictors organized across seven clinical categories, spanning demographic details, allergic history, respiratory symptoms, infections, medication exposure, sleep patterns, and family background. The children were then followed for a median of 4.5 years, until they reached school age of six years or older, at which point asthma status was established. This prospective design matters, because most existing asthma prediction tools were built retrospectively or in general birth cohorts rather than in a population of food-allergic infants, a group whose risk profile differs substantially from that of the general pediatric population.

At the heart of the work is a two-stage, performance-weighted ensemble that combines four very different families of algorithms: Random Forest, XGBoost, LightGBM, and a deep residual network. Each base learner captures different structure in the data. Tree-based gradient boosting methods such as XGBoost and LightGBM excel at nonlinear interactions among tabular clinical variables, Random Forests provide variance reduction through bagging, and the deep residual network can model higher-order feature relationships. Rather than averaging their outputs naively, the researchers assigned fusion weights proportional to each model’s measured performance, allowing stronger learners to dominate the final risk score while weaker ones still contribute complementary signals.

The most distinctive technical choice, however, is the cost-sensitive optimization. In asthma screening, the two types of error are not equivalent: a false alarm means extra monitoring for a child who may never develop the disease, but a missed case means a high-risk infant slips through without preventive attention, potentially delaying diagnosis and treatment that studies have linked to worse lung function outcomes. The team therefore trained the ensemble with a missed-to-false-alarm cost ratio of 5:1, deliberately penalizing missed cases five times more heavily. The result, on an independent test set, was an area under the receiver operating characteristic curve of 0.886 and a recall of 1.00, meaning that not a single child who went on to develop asthma was missed by the model at the chosen threshold.

Sensitivity analyses added an important nuance: the zero-missed performance remained stable across cost ratios of 5:1 and above, suggesting that the 5:1 setting was not itself the key driver of the perfect recall. Instead, the ensemble’s underlying discrimination appears strong enough that a range of asymmetric cost settings yields the same reassuring property. The researchers also evaluated clinical utility beyond raw discrimination using decision-curve analysis, a method that quantifies the net benefit of acting on a model’s predictions across a range of threshold probabilities. The model showed superior net benefit across broad threshold ranges, with an optimal operating cutoff identified at a predicted probability of just 0.09, a low threshold consistent with the deliberately sensitive screening posture.

Interpretability was addressed through SHAP-based feature importance analysis, which ranked the predictors driving individual predictions. The top contributors were respiratory syncytial virus infection, the frequency of wheezing episodes, characteristics of nocturnal sleep, cumulative antibiotic exposure, and the duration of rhinitis. Several of these echo long-standing clinical intuition, particularly wheeze grading, which has anchored classical tools such as the asthma predictive index for two decades. Others are more provocative. The prominence of antibiotic exposure aligns with a growing body of literature linking early-life antibiotics, possibly through gut microbiome disruption, to later allergic disease, although the observational nature of the data means the model captures association rather than proven causation.

The team also engineered three clinically interpretable features from the raw variables before model training, a step intended to translate pediatric knowledge into representations the algorithms could exploit more efficiently. Interaction analysis of the top ten feature pairs, visualized alongside the SHAP results, suggests that combinations such as wheeze frequency with infection history carry predictive information beyond what either variable contributes alone. Radar charts comparing three threshold strategies and calibration curves assessing the agreement between predicted probabilities and observed outcomes round out a performance evaluation that is unusually thorough for a clinical prediction study of this size.

The clinical logic behind the effort is straightforward. Delayed asthma diagnosis has been associated with more frequent exacerbations and measurable decrements in pulmonary function, while early intervention with inhaled corticosteroids in persistent disease has documented benefits. A screening tool that works before age two, using nothing more than clinical history, could allow primary care providers to stratify food-allergic infants into intensive follow-up versus routine monitoring, individualizing surveillance long before symptoms become unmistakable. Because the model requires no blood tests, lung function measurements, or imaging, the barrier to deployment in community settings is comparatively low, at least in principle.

Important caveats remain. The cohort of 447 children came from a single Chinese hospital network, and the authors themselves describe the model as a preliminary predictor, noting that external validation in independent, geographically diverse populations will be essential before any clinical adoption. Individual-level data are not publicly available for privacy reasons, which limits independent scrutiny, although aggregated results and supplementary tables are provided. Calibration, threshold selection, and the stability of the top predictors all need to be confirmed in settings with different healthcare patterns, antibiotic prescribing habits, and RSV epidemiology. The 1.00 recall figure, while striking, will also need replication; perfect sensitivity on a single test set can reflect favorable data splits as well as genuine generalizable performance.

Even with those qualifications, the study represents a meaningful step toward proactive, data-driven allergy care. It demonstrates that a carefully engineered ensemble, tuned for the asymmetric costs that actually matter to families and clinicians, can extract a usable school-age asthma risk signal from information gathered in the first two years of life. If validated externally, the approach could shift the management of food-allergic infants from watchful waiting toward targeted prevention, giving the children at highest risk a head start on the interventions that preserve lung function and, ideally, change the trajectory of the allergic march before it hardens into chronic disease.

Subject of Research: Machine-learning prediction of school-age asthma risk in infants with food allergy

Article Title: Ensemble model–based prediction of school–age asthma risk in children with food allergy

Article References: Ensemble model–based prediction of school–age asthma risk in children with food allergy. (n.d.). https://doi.org/10.1038/s41390-026-05553-9

Image Credits: AI Generated

DOI: 10.1038/s41390-026-05553-9

Keywords: food allergy, asthma prediction, machine learning, ensemble model, pediatrics, cost-sensitive learning, SHAP, RSV infection, wheezing, antibiotic exposure, prospective cohort, early screening