February 2026
Eosinophils in the Real World: Spotting Hidden Hypereosinophilic Syndrome with Prediction Modeling
Article: Prediction model to identify patients with hypereosinophilic syndrome using real-world data
Khoury P, Chung Y, Carsten D, et al.
J Allergy Clin Immunol Glob. 2025
Reviewed by Stella Oyewole, University of Cincinnati College of Medicine, Ohio, United States
Hypereosinophilic syndrome (HES) is a rare and heterogeneous disorder defined by persistent peripheral eosinophilia (≥1,500/µL) accompanied by tissue or organ damage. The diagnosis of HES can be challenging owing to variable clinical presentations and overlap with other eosinophilic disorders, thus may be underrecognized. Machine learning predictive tools may offer an approach to better understand HES prevalence. Read more
In this study, investigators developed a machine learning–based prediction model using an open claims database including patients (with/without HES ICD-10 code) who had ≥2 blood eosinophil counts >1,000 cells/μL. This lower threshold was chosen to capture a broader population, including patients with treatment-suppressed counts. Selected predictors included demographics, disease manifestations, comorbidities, treatments, and healthcare utilization.
Selected variables were incorporated into a statistical model to estimate predicted probabilities of HES risk across thresholds from 0.1–0.9. Key predictors included having a bone marrow biopsy, blood eosinophil count >3,000/µL, and history of asthma. Model performance was evaluated using Receiver Operating Characteristics Area under the curve (ROC AUC) and Precision-Recall AUC (PR AUC) to assess overall discrimination and precision, both demonstrated strong results (ROC AUC = 0.82; PR AUC = 0.83). At a 0.7 probability threshold, 6,233 additional patients with predicted HES were identified and exhibited characteristics similar to diagnosed cases. The estimated model-derived prevalence of total HES (diagnosed plus predicted) was 5.65% among individuals with elevated eosinophil counts, suggesting underdiagnosis.
Strengths of this study include a large real-world dataset and a robust predictive approach. Limitations include potential misclassification of HES in secondary datasets and limited generalizability to patients with lower blood eosinophil counts. Given that HES is often diagnosed late, this model may enable earlier identification of HES patients, prompting timely referral & intervention, although it is not a stand-alone diagnostic tool and requires clinical interpretation and further validation. Rather than distinguishing primary (clonal/idiopathic) HES from secondary (reactive) eosinophilia, the model is designed to flag patients whose overall patter of findings resemble those with a recorded HES diagnosis. In addition to external validation of model and extension to identification of tissue-based HES, future studies could be focused on developing machine-learning based tools to facilitate earlier diagnosis of HES in the population. Overall, these findings suggest that machine learning-based prediction may improve HES detection and reveal an underrecognized disease burden.
Stella Oyewole is a PhD candidate at the University of Cincinnati College of Medicine. She earned her Doctor of Dental Surgery (DDS) degree from Obafemi Awolowo University in Nigeria. Her doctoral research focuses on evaluating cell- and tissue-informed circulating biomarkers to assess tissue-specific disease activity in eosinophil-associated diseases (EAD). Under the mentorship of Dr. Nives Zimmermann, her work aims to support earlier and more precise diagnosis of these disorders and improve understanding of their underlying biology.