If you can predict which itchy toddler skin cases turn into school-age asthma, you can watch the right kids earlier, which means you might intervene sooner, which could make the whole allergic domino chain a little less rude. That is the basic pitch of this 2026 paper - and honestly, it is a strong one. Kids with early atopic dermatitis, or eczema, do not all follow the so-called atopic march from skin trouble to hay fever to asthma. Some do. Some do not. Medicine has been stuck doing a lot of educated shrugging. This study tries to replace some of that shrugging with numbers.
Researchers used longitudinal electronic health record data from Kaiser Permanente Southern California to study 10,688 children diagnosed with atopic dermatitis before age 3.[^1] They built two machine learning models for each later outcome: asthma and allergic rhinitis. One model used a fuller pile of structured EHR variables, and the other used a simplified set of more routine clinical features.
The asthma results were very good. The comprehensive model reached an AUC of 0.893, and the simplified one was basically tied at 0.892.[^1] That is impressive, partly because simplified models are the clinical equivalent of packing light and still arriving with your socks matched. The allergic rhinitis models were less flashy but still useful, with AUCs of 0.793 and 0.773.[^1]
On one hand, that means early-life records really do contain signal about who is likely to develop moderate-to-severe persistent asthma or rhinitis by ages 5 to 11. On the other hand, this is not a crystal ball with a tiny stethoscope. At high specificity, sensitivity stayed modest. In plain English: the model is pretty good at identifying some high-risk kids without flagging too many others, but it will still miss plenty. That matters.
Why this is interesting in a mildly unsettling way
Atopic dermatitis has long been known as a possible opening act for later allergic disease. The general pattern is often called the atopic march, though real life is messier than the diagram in your immunology textbook nobody proofread.[^2][^3] Some children start with eczema and later develop allergic rhinitis or asthma. Others do not. Clinicians would very much like to know which child is which before the wheezing starts at 2 a.m. and everyone’s weekend gets cancelled by inflammation.
That is where machine learning earns its keep. Not by being magical, but by being annoyingly good at spotting patterns across piles of messy records that no human could hold in their head at once. Think of it as giving the overworked pediatric chart a second job as a fortune teller, except this fortune teller speaks in risk strata and calibration plots instead of incense smoke.
Recent reviews suggest this fits a broader trend. A 2025 systematic review and meta-analysis found that machine learning has been widely used to identify asthma and allergy trajectories in children, but also noted major variation in methods and a need for stronger validation before real clinical deployment.[^4] A 2024 scoping review of pediatric asthma ML studies found promise across prediction, diagnosis, and phenotype classification, while repeatedly running into the same trio of headaches: data quality, small samples, and interpretability.[^5] Welcome to medical AI, where the ambition is huge and the fine print is doing cardio.
What this could change - if it holds up
If these models reproduce in other health systems, they could help pediatricians decide which children with early eczema deserve closer follow-up, earlier allergy workups, more aggressive prevention counseling, or targeted monitoring. That is not trivial. Asthma is one of those diseases that can sit quietly until it absolutely does not. Allergic rhinitis sounds gentler, but in many kids it is part of the same inflammatory road network.
There is also a practical point here: the simplified model performed almost as well as the comprehensive one for asthma.[^1] That makes deployment feel more realistic. Hospitals love “state of the art” right up until the model requires seventeen custom variables, six missing-data rituals, and a moon phase. A simpler model has a better chance of escaping the PDF and entering actual clinic workflow.
Still, caution first. This was a retrospective cohort from one integrated care system.[^1] EHR-based models can inherit the biases of who gets diagnosed, who gets follow-up, and who gets documented thoroughly enough to exist in the data as more than a billing event with a pulse. Broader reviews of fair machine learning in real-world health data make exactly this point: performance is only half the story if the model distributes errors unevenly across populations.[^6]
So yes, this paper is hopeful. It suggests early eczema might be more than a warning sign - it might be a prediction window. But it also leaves you with that distinctly 2026 feeling: on one hand, we are getting better at seeing risk before disease fully announces itself. On the other hand, we are trusting giant clinical databases and statistical machinery to help decide which child needs more attention. Useful? Absolutely. A little uncanny? Also absolutely. Possibly both. Definitely both.
References
[^1]: Chen W, Zhou B, Schatz M, et al. Machine learning prediction of asthma and allergic rhinitis in children with early-onset atopic dermatitis. J Allergy Clin Immunol. Published online April 17, 2026. doi:10.1016/j.jaci.2026.03.025. PubMed: https://pubmed.ncbi.nlm.nih.gov/42002051/
[^2]: DiGiacomo D. From Eczema to Allergies to Asthma: Understanding the Allergic March. HealthyChildren.org. 2024. https://www.healthychildren.org/English/health-issues/conditions/allergies-asthma/Pages/from-eczema-to-allergies-to-asthma-understanding-the-allergic-march.aspx
[^3]: Hill DA, Spergel JM. The Atopic March: Critical Evidence and Clinical Relevance. Ann Allergy Asthma Immunol. 2018;120(2):131-137. doi:10.1016/j.anai.2017.10.037. PMCID: PMC5806141
[^4]: Lisik D, Özuygur Ermis SS, Milani GPI, et al. Machine learning-derived asthma and allergy trajectories in children: a systematic review and meta-analysis. Eur Respir Rev. 2025. doi:10.1183/16000617.0160-2024. PubMed: https://pubmed.ncbi.nlm.nih.gov/39778923/
[^5]: Ojha T, Patel A, Sivapragasam K, et al. Exploring Machine Learning Applications in Pediatric Asthma Management: Scoping Review. JMIR AI. 2024;3:e57983. doi:10.2196/57983
[^6]: Huang Y, Guo J, Chen WH, et al. A scoping review of fair machine learning techniques when using real-world data. J Biomed Inform. 2024;151:104622. doi:10.1016/j.jbi.2024.104622
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.