Machine Learning Classifiers Identify Atopic Dermatitis Patterns in Children and Adults

In a retrospective cross-sectional analysis of multiracial population data from the United States, researchers developed machine learning models capable of distinguishing atopic dermatitis cases in both pediatric and adult cohorts with high discriminatory accuracy. The investigation evaluated data from the National Health and Nutrition Examination Survey spanning the 1999 through 2006 cycles, focusing on individuals who completed dermatology and allergy questionnaires. By implementing dedicated class balancing protocols to mitigate the severe skew inherent in population-level surveys, the investigators addressed a long-standing obstacle in clinical dermatology informatics: the persistent scarcity of validated, adult-specific diagnostic classification tools alongside childhood models. Beyond achieving balanced discrimination across age groups, the application of explainable artificial intelligence frameworks exposed contrasting physiological relationships, most notably demonstrating that composite dietary antioxidant intake exerts opposing associations with disease probability in children compared to mature adults.

Atopic dermatitis represents a chronic, pruritic inflammatory skin disorder that affects approximately 204.05 million individuals across the globe, presenting an immense burden on physical comfort, mental health, and healthcare delivery systems. In the United States, twelve-month diagnosed prevalence rates reach approximately 15.8 to 17.0 percent among children between six months and eighteen years of age, alongside a 7.3 percent prevalence rate in adult populations. Contemporary medical management relies heavily on therapeutic interventions including topical corticosteroids, calcineurin inhibitors, phototherapy, conventional systemic immunosuppressants, Janus kinase inhibitors, and biologic therapies. However, these clinical interventions can carry non-trivial clinical risks, encompassing secondary cutaneous infections, blepharitis, conjunctivitis, head and neck dermatitis, exacerbations of acne, and potential malignancies. Furthermore, the disease frequently initiates the atopic march, progressing toward subsequent allergic conditions such as asthma and allergic rhinitis, making early and precise characterization essential for optimizing patient outcomes.

Developing automated clinical classifiers has historically been confounded by substantial pathophysiological heterogeneity between early-onset childhood manifestations and adult-onset presentations. While pediatric atopic dermatitis has received considerable attention in predictive research, adult populations demonstrate distinct cutaneous lesion distributions, mature immune cytokine environments, and complex accumulations of environmental, occupational, and lifestyle exposures. Prior modeling attempts frequently suffered from suboptimal diagnostic accuracy, an absence of interpretable decision-support frameworks for treating clinicians, and severe dataset class imbalance that caused classifiers to default to majority non-disease predictions. When classifiers encounter datasets where unaffected individuals vastly outnumber diagnosed cases, conventional learning algorithms frequently report misleadingly high overall accuracy while failing to identify true positive cases, generating an excessive burden of false negative results that undermines utility in clinical screening environments.

To overcome these systemic analytical hurdles, the research team curated records from 41,474 total survey respondents across four consecutive two-year survey cycles of the National Health and Nutrition Examination Survey. Following the exclusion of 2,274 participants who were either pregnant or reported a medical history of cancer, the final analytical cohort comprised 22,635 eligible individuals. These participants were partitioned by biological maturity into 8,297 pediatric subjects under twenty years of age and 14,338 adult subjects aged twenty years or older. Case definitions were established using standardized household survey items, defining atopic dermatitis through self-reported affirmative answers to recent past-year dermatitis or rash symptoms, concurrent active rashes, chronic six-month itchy rash histories, or direct clinical notifications from medical doctors confirming an eczema diagnosis. Among qualified participants, 1,118 children, representing 13.47 percent, and 1,732 adults, representing 12.80 percent, met the criteria for atopic dermatitis.

The researchers extracted 1,175 raw variables spanning demographic interviews, dietary consumption recalls, clinical physical examinations, laboratory assays, and health questionnaires, subsequently unifying and reconstructing them into forty-four clinically accessible predictive features. These candidate features captured vital socioeconomic indicators including the poverty income ratio, educational attainment, household income, race, marital status, and household size, alongside biochemical and hematological indices such as C-reactive protein, the systemic immunity-inflammation index, the systemic immune response index, and the nutritional risk index. Comorbid chronic conditions, including asthma, hay fever, chronic bronchitis, emphysema, depression, thyroid disease, diabetes, and cardiovascular disorders, were simultaneously cataloged. To maintain methodological transparency, variables exhibiting over eighty percent missing data were excluded, while remaining missing numerical and categorical values were imputed deterministically using cohort means and modes rather than stochastic synthesis.

Recognizing that data preprocessing can inadvertently introduce bias or data leakage, the investigators split the cohort into training and testing partitions prior to executing any normalization, imputation, or class balancing routines. The raw baseline models, trained without balancing interventions, suffered dramatically from the roughly one-to-seven class imbalance, yielding poor accuracy scores between 0.500 and 0.503 and misclassifying nearly all positive cases. To resolve this deficiency, the researchers systematically tested three distinct balancing paradigms across ten machine learning architectures: oversampling via synthetic minority techniques, ensemble sampling using cleaning algorithms such as repeated edited nearest neighbors combined with boundary-focused oversampling, and controlled under-sampling. In total, the team constructed and tuned 820 separate model iterations, evaluating algorithms such as Random Forest, Light Gradient Boosting Machine, Support Vector Machines, Gradient Boosting Trees, and Categorical Boosting.

When assessed through internal five-fold cross-validation and rigorous internal hold-out testing on unseen raw samples, the under-sampling approach, designated Model-US, demonstrated the most effective discriminatory performance. In childhood atopic dermatitis, the optimal Random Forest architecture achieved an overall accuracy of 0.885, accompanied by an area under the receiver operating characteristic curve of 0.957, an F1-score of 0.884, a sensitivity recall of 0.877, and a precision of 0.892. This refinement dramatically slashed false negative counts from 275 in baseline oversampling comparisons down to 37 cases in the pediatric testing subset. In adult atopic dermatitis, the Light Gradient Boosting Machine yielded a classification accuracy of 0.835, an area under the curve of 0.913, and curtailed false negatives from 276 down to 51 cases, confirming that reducing majority-class noise allowed decision trees to capture genuine biological and demographic patterns.

To demystify the internal decision logic of the optimized classifiers, the authors implemented Permutation Feature Importance and SHapley Additive exPlanations, two established model-agnostic interpretability methods. In the childhood model, both analytical frameworks identified an identical hierarchy among the top seven predictive features: the birthplace of adult household reference persons, the poverty income ratio, race, personal asthma history, patient age, body mass index, and the composite dietary antioxidant index. Adult family reference members born in Mexico and children identifying as Mexican American exhibited strong, consistent negative associations with atopic dermatitis probability, whereas household reference persons born within the United States or other foreign regions correlated with elevated disease classification probabilities. Asthma history, younger age, female sex, lower body mass index, and elevated poverty income ratios further amplified childhood disease associations.

Adult atopic dermatitis classifications demonstrated an eighty percent overlap in the top fifteen features ranked by permutation importance and Shapley values, highlighting participant birthplace, race, poverty income ratio, antioxidant consumption, inflammatory indices, and respiratory conditions. Similar to the pediatric cohort, adult participants born in Mexico displayed substantial negative associations with disease classification, while individuals identifying as non-Hispanic White exhibited the strongest positive attribution toward an atopic dermatitis designation. Furthermore, individual sample force analyses demonstrated that personal histories of depression and chronic bronchitis exerted the most intense positive impacts on adult case classifications, followed by contributory roles from arthritis, thyroid disorders, diabetes, history of military service, and stroke. Elevations in the systemic immune response index, systemic immunity-inflammation index, and C-reactive protein consistently skewed predictions toward positive disease status.

One of the most biologically intriguing insights generated through model interpretation involved the divergent role of dietary antioxidant consumption across different life stages. Evaluating nutritional recalls using the Composite Dietary Antioxidant Index—calculated initially across vitamin C, zinc, and selenium, and subsequently expanded to include vitamin A, vitamin E, and beta-carotene—revealed reciprocal relationships. In pediatric subjects, elevated dietary antioxidant consumption correlated significantly with negative Shapley values, indicating a protective association against atopic dermatitis classification. Conversely, in mature adults, higher dietary antioxidant intake yielded positive Shapley values, aligning with an increased probability of atopic dermatitis assignment. This contrast aligns with prior meta-analyses demonstrating that clinical antioxidant supplementation confers measurable symptom reductions exclusively in children, reflecting age-dependent shifts in immune cytokine profiles and mature barrier pathophysiology.

The investigators emphasized several structural constraints that govern the clinical interpretation of their retrospective findings. Foremost, cross-sectional survey architectures inherently capture statistical associations rather than direct biological causality, leaving unresolved whether nutritional intakes or inflammatory markers actively precipitate or merely reflect underlying atopic processes. Additionally, outcome classifications relied entirely on self-reported questionnaire responses rather than prospective clinical examinations, objective morphological evaluations, or cutaneous biopsies, introducing susceptibility to diagnostic recall errors. Furthermore, the internal hold-out validation was conducted upon balanced test subsets of three hundred cases and three hundred non-cases to prevent majority-class masking, meaning real-world screening precision will adjust downward in general populations where clinical prevalence ranges between seven and seventeen percent.

Finally, the researchers noted that reliance on survey cycles from 1999 to 2006, while necessary to capture uniform dermatology questionnaires, precedes subsequent two-decade shifts in environmental pollution, ultra-processed dietary habits, and evolving clinical diagnostic standards. Moreover, deterministic mean and mode imputations, though adopted to avoid complex unverified stochastic modeling assumptions, inevitably compress variable variance. As the study omitted formal decision curve analyses and calibration curves, the generated risk estimates represent relative classification probabilities rather than calibrated absolute diagnostic thresholds. Nonetheless, the development of dual-cohort machine learning classifiers offers an empirical foundation for future primary care screening tools, providing clinicians with interpretable, age-tailored frameworks to identify individuals facing elevated probabilities of atopic skin disease.

Subject of Research: Machine learning classification and explainable artificial intelligence profiling of childhood and adult atopic dermatitis

Article Title: Clinically actionable classification tool for childhood and adult atopic dermatitis via explainable AI: a retrospective cross-sectional study optimizing diagnostic accuracy

Article References: Zhou, R., Zhang, W., Zhou, L., Zhu, Q., Zhang, J., Wang, S., & Wang, C. (2026). Clinically actionable classification tool for childhood and adult atopic dermatitis via explainable AI: a retrospective cross-sectional study optimizing diagnostic accuracy. BioData Mining, 19(1), Article 66. https://doi.org/10.1186/s13040-026-00565-x

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00565-x

Keywords: atopic dermatitis, machine learning, explainable artificial intelligence, NHANES, dermatology, pediatric health, adult dermatology, dietary antioxidants, clinical classification, Clinically, actionable, classification