Model Signals Primary Immunodeficiency Disease Based on History of Symptomatic Care
A machine learning model trained using clinical data from electronic health records (EHR) can aid in the early diagnosis of patients with primary immunodeficiency disease (PID) based on their history of symptomatic care, according to a study published online in The Journal of Allergy and Clinical Immunology: In Practice.
“Our model may serve as an essential first step to the development of an EHR alert for practitioners to consider a diagnosis of primary immunodeficiency disease,” wrote corresponding author Christina E. Ciaccio, MD, of the University of Chicago Department of Pediatrics, and coauthors, “thereby expediating treatment using immunoglobulin replacement therapy, potentially reducing risk of morbidity and mortality.”
Researchers conducted a retrospective study of University of Chicago Medical Center patients with PID and a control group of patients with asthma matched for age, sex, and race. Potential predictors of PID focused on comorbidities, laboratory tests, medications, and radiological orders that were indicative of symptom-related treatment before diagnosis. Among a total 6422 patients in the cohort, 247 were diagnosed with primary immunodeficiency disease.
The logistic regression model based on comorbidities demonstrated good discrimination between patients with PID and patients with asthma, with an area under the curve of 0.62, according to the study. However, adding laboratory results, medications, and radiological orders improved the discrimination performance, with an area under the curve of 0.70.
“We illustrate how including laboratory tests, radiological orders, and medications from a patient’s historical interaction with a hospital increases the performance of detecting primary immunodeficiency disease as opposed to just using International Classification of Diseases-based comorbidities, in terms of better discrimination as well as increased sensitivity and specificity,” researchers wrote. “Notably, the number of radiological procedures ordered was the most important prediction.”
Extending from logistic regression to more advanced machine learning frameworks did not enhance the model’s performance, researchers reported.
Reference
Mayampurath A, Ajith A, Anderson-Smits C, et al. Early diagnosis of primary immunodeficiency disease using clinical data and machine learning. J Allergy Clin Immunol Pract. 2022;10(11):3002-3007.e5. doi: DOI: 10.1016/j.jaip.2022.08.041
