A predictive model for identification of pediatric individuals with common variable immunodeficiency through electronic medical records

Nouf Alsaati; Chris Penney; Ingo Helbig; Kathleen E Sullivan

doi:10.1016/j.jaci.2025.02.032

A predictive model for identification of pediatric individuals with common variable immunodeficiency through electronic medical records

J Allergy Clin Immunol. 2025 Mar 7:S0091-6749(25)00263-5. doi: 10.1016/j.jaci.2025.02.032. Online ahead of print.

Authors

Nouf Alsaati¹, Chris Penney², Ingo Helbig³, Kathleen E Sullivan⁴

Affiliations

¹ Division of Allergy and Immunology, Children's Hospital of Philadelphia, Philadelphia, Pa.
² Department of Biomedical and Health Informatics, Children's Hospital of Philadelphia, Philadelphia, Pa.
³ Division of Neurology, Epilepsy NeuroGenetics Initiative (ENGIN), Children's Hospital of Philadelphia, Philadelphia, Pa; Department of Neurology, University of Pennsylvania, Perelman School of Medicine, Philadelphia, Pa.
⁴ Division of Allergy and Immunology, Children's Hospital of Philadelphia, Philadelphia, Pa. Electronic address: sullivank@chop.edu.

PMID: 40058414
DOI: 10.1016/j.jaci.2025.02.032

Abstract

Introduction: Common variable immunodeficiency (CVID) is characterized by recurrent sinopulmonary infections. However, in the pediatric population, recurrent sinopulmonary infections early in life are common, which can render key clinical features of CVID less distinctive. Accordingly, the diagnosis of CVID is often delayed owing to the heterogeneous nature of the presentation and the broad range of ages of onset. A 10-year lag in diagnosis has been found for CVID, and there is a critical need for improved time to diagnosis.

Objective: Our aim was to utilize machine learning techniques to identify a clinical signature of CVID in a pediatric population.

Methods: Our selected cohort included 112 individuals with CVID and 627 controls. The controls were restricted from having other medical conditions associated with infection. A machine learning data set was constructed by summing patient-level counts of clinical metrics. A total of 3 supervised machine learning classifiers were trained, tuned, and performance-tested. We validated our findings using a distinct control cohort with high medical complexity and tested a logistic regression approach.

Results: Key features associated with CVID were chest radiograph count, number of antibiotic prescriptions, and number of common infections. Our Extreme Gradient Boosting (XGBoost) model best predicted eventual CVID diagnosis, with an F1 score of 0.77, a total of 21 of 29 CVID diagnoses classified correctly (8 false-negative results), and 179 of 183 patients without CVID correctly classified (4 false-positive results) up to 10 years before the eventual clinical diagnosis. Key features with a robust association with pediatric CVID were the frequency of common infections and antibiotic prescriptions.

Conclusion: In spite of a high frequency of infections in the comparator population, the clinical signature of pediatric CVID was sufficiently distinctive to enable early identification.

Keywords: Primary immunodeficiency; antibody deficiency; artificial intelligence; inborn error of immunity; machine learning; predictive algorithm.