A machine-learning model built from six clinical measures performed better than three other approaches at identifying ADHD comorbidity among children with allergic rhinitis, according to a retrospective cohort study. On an independent test portion of the same cohort, the XGBoost model recorded an area under the receiver operating characteristic curve, or AUC, of 0.878. AUC is a measure of how well a model separates children with and without the condition across possible decision cutoffs.
At the study’s selected probability cutoff of 0.527, the model correctly identified 77.8% of the children with ADHD and correctly ruled it out in 83.1% of those without it. Overall accuracy was 80.6%, with an F1-score of 0.789. The comparison AUCs were 0.814 for a support-vector machine, 0.811 for a random forest and 0.713 for logistic regression.
What the model used
The cohort included 518 children with allergic rhinitis: 255 had ADHD comorbidity and 263 were controls without ADHD. The researchers divided the group randomly into a 70% training set and a 30% test set. They used the training data to select variables and build the models, leaving the test group for the reported internal assessment.
A hybrid selection process retained Total IgE, Vitamin D, body mass index, ferritin, sleep duration and screen time. The selection was carried out only in the training set, using LASSO with 10-fold cross-validation, Boruta and a random-forest feature ranking method. This design was intended to reduce leakage of information from the test data into model building.
The study used a retrospective allergic-rhinitis cohort in which ADHD status was confirmed with DSM-5 criteria. Children with severe congenital anomalies, other major psychiatric disorders, severe missing data or severe or uncontrolled concurrent allergic disease were excluded. The methods report ethics approval and written consent from parents or guardians.
IgE and Vitamin D stood out
The model’s explanation tools pointed most strongly to Total IgE, body mass index and Vitamin D. Their mean absolute SHAP values, a model-based measure of how much each variable contributed to individual predictions, were 0.981, 0.854 and 0.601 respectively. The strongest reported interaction was between Total IgE and Vitamin D, with an interaction strength of 0.20.
The patterns were not simply linear. The model’s SHAP analysis showed an IgE inflection at about 250 to 300 IU/mL and a steep Vitamin D pattern below 20 ng/mL. Its joint-risk map reached its maximum contribution above 500 IU/mL of IgE combined with Vitamin D below 20 ng/mL. In a four-group comparison using IgE above 250 IU/mL and Vitamin D below 20 ng/mL, the dual-high-risk group had a greater model contribution than the other groups, with p below 0.001.
The underlying group differences pointed in the same direction: children with ADHD had higher Total IgE than controls, 251.00 versus 149.00 IU/mL, and lower Vitamin D, 24.17 versus 26.32 ng/mL. Both comparisons had p-values below 0.001. These findings describe associations in the cohort, not proof that IgE or Vitamin D causes ADHD.
A possible triage aid, not a diagnosis
The model also separated the test cohort into three predicted risk groups. Observed ADHD prevalence was 18.3% in the low-risk group and 73.4% in the high-risk group, with a statistically significant trend across the groups at p below 0.001. Decision-curve analysis found greater net benefit for the model than for treating all children or treating none across threshold probabilities from 0.01 to 0.65.
Its predicted probabilities appeared aligned with observed ADHD prevalence in the test group, and a Hosmer-Lemeshow calibration test was non-significant at p = 0.315. Calibration here concerns how closely predicted and observed rates matched in this test cohort, not whether the model will perform the same way elsewhere.
A separate adjusted logistic-regression analysis reported higher odds associated with Total IgE and screen time, with odds ratios of 1.87 and 1.51, and lower odds associated with Vitamin D and sleep duration, with odds ratios of 0.61 and 0.71. The reported 95% confidence intervals were 1.45 to 2.49 for IgE, 1.19 to 1.93 for screen time, 0.48 to 0.77 for Vitamin D and 0.56 to 0.89 for sleep duration. These are retrospective associations, not evidence of cause and effect.
The researchers describe the model as a possible tool for stratifying children for further assessment. But the test set was a random holdout from the same retrospective cohort, and the supplied analysis reports no prospective or external validation of discrimination, calibration or clinical utility. That means the figures show how the model performed within this dataset, not how it will perform in other pediatric settings.
Nor do these associations establish that changing IgE, Vitamin D, sleep or screen time would prevent or treat ADHD. The thresholds in this analysis were exploratory, and their stability outside this cohort remains untested. Prospective multicenter studies will need to assess whether the model and its apparent IgE and Vitamin D patterns hold across pediatric populations and treatment settings.
The authors declared that they received no financial support for the work or its publication and reported no commercial or financial relationships that could be construed as a potential conflict of interest. They stated that raw data supporting the conclusions will be made available without undue reservation.
Paper data and sources
Original title: Development and validation of an interpretable machine learning model for predicting ADHD comorbidity in children with allergic rhinitis: a SHAP-based analysis
Authors: Yuan H, Wu N
Journal/Repository: Frontiers in public health
Status: Peer-reviewed
First online: 2026-10-08
DOI: Not available
Original paper · Full text