Trajectory Analysis of Pediatric Longitudinal Allergen Sensitization Patterns From Real World Data.

purposeTo explore the relationship between allergen sensitization patterns and allergic disease diagnosis, with a particular focus on longitudinal sensitization trajectories. Using a large pediatric and young adult cohort from a common data model (CDM)-formatted clinical database, we applied unsupervised learning techniques to identify both static and dynamic sensitization clusters. We developed a predictive model to assess how early sensitization profiles contribute to developing allergic disease later in life.

methodsWe conducted a single-center, retrospective cohort study using electronic health records from the Seoul National University Hospital CDM. Patients who underwent a skin prick test, multiple allergen simultaneous test, or ImmunoCAP between 2002 and 2021 were included if they were tested in 2 or more age groups. Seventeen allergen categories were analyzed. Latent class analysis identified cross-sectional sensitization clusters, and Gaussian mixture models were used to derive longitudinal trajectories. A Categorical Boosting multilabel classification model was used to predict allergic diagnoses with interpretability enhanced by SHapley Additive exPlanations.

resultsA total of 1,656 patients were analyzed. Four cross-sectional clusters were identified (house dust mite [HDM]-dominant, HDM + food, polysensitization, and multiple inhalant clusters). Nine longitudinal trajectories showed a developmental shift from food sensitization to aeroallergen sensitization. Early HDM sensitization was strongly associated with subsequent asthma, whereas eczema and food allergies are linked to food or polysensitization trajectories. Rhinitis was common across aeroallergen clusters. The prediction performance was robust for asthma (F1 score, 0.73; area under the curve, 0.80), moderate for eczema and food allergies, and limited for rhinitis.

conclusionsLongitudinal sensitization trajectories, particularly early HDM dominance, were predictive of asthma development, whereas the predictive values for eczema, food allergies, and rhinitis were weaker. Modeling sensitization history using machine learning may enable the early identification of high-risk children and inform personalized strategies for allergy prevention.