Predicting Food Allergies & Eczema: A Practical Machine-Learning Empowered Approach

by Michael Brandwein

Coauthors: Kirill Pevzner, Michal Ashkenazi, Michael Brandwein MA, KP, & MB report personal fees from MYOR Diagnostics Ltd., during the conduct of the study; In addition, MB has patent US 2010/0285110 pending

Medical Devices & Digital Health

Background: The rise in prevalence of atopic dermatitis (AD) has been correlated with numerous elements of the exposome and modern-day lifestyle. Parental history of AD and other atopic conditions are additional risk factors for disease development. We posited that the combined analysis of familial history and other risk elements may allow us to understand the interaction between the two, and the driving factors behind an infant’s risk of developing AD. We employed various prediction models to assess an infant’s risk of developing AD from birth using a geographically diverse cohort and a host of easily-assessed risk factors.
Methods: We combined study data from four separate pediatric cohorts. Each cohort assessed the development of AD within 15 months of birth, and gathered data on an infant’s paternal and maternal history of AD, asthma and food allergies. Additional documented risk factors include birthweight, gender, gestational age at birth, household pets, mode of delivery, maternal smoking and ethnicity. Predictive models were trained and validated on the combined dataset.
Results: The odds ratio conferred on an infant due to paternal history of AD (3.585), food allergies (3.121) or asthma (3.134) is higher than that of maternal AD (2.434), food allergies (2.156) and asthma (1.973). Receiver operating characteristic curve analyses showed an area under the curve of 0.83 for the random forest model, which outperformed a logistic regression and XGBoost model. Maximum accuracy of 78%, with corresponding sensitivity of 63%, specificity of 87%, and a positive predictive value of 75% was obtained. Birthweight, gestational age at birth, household pets and parental history of AD are the most important factors contributing towards risk prediction.
Conclusion: Machine learning predictive modeling using non-invasive and accessible inputs can serve as a powerful tool to stratify an infant’s risk of developing atopic dermatitis from birth. Knowledge of an infant’s risk can inform both caregivers and medical professionals as to timely interventions to mitigate the discomfort associated with AD.