Artificial Intelligence in Allergy and Immunology: Comparing Risk Prediction Models to Help Screen Inborn Errors of Immunity

To develop and compare risk prediction models, using logistic regression and machine learning (ML) techniques, to measure the individual chance of a confirmed diagnosis of inborn errors of immunity (IEI) in children at risk for this problem, according to previous evaluation by general pediatricians and clinicians.

This study is a retrospective longitudinal cohort study of 128 subjects with suspected IEI from a pediatric immunology outpatient clinic of a state university in Brazil for the year 2018. Inclusion criteria for laboratory included at least a complete blood cell count, serum immunoglobulins, C3, C4, rubella, and antihepatitis B surface antibody serology, and HIV serology, or viral load for children under age 2 years. Exclusion criteria included positive HIV serology or viral load.

At visit 1, patients had a comprehensive medical history, physical examination, and laboratory tests done. After analysis of the above, patients were classified into 2 subgroups: subgroup 0 (control) – no clinical or laboratory alteration found, and subgroup 1 (IEI risk), when IEI was highly suspicious or confirmed. A combined risk score was used based on the sum of risk for any primary IEI criteria (given 1 point each) listed in the Jeffrey Modell Foundation 10 Warning Sings for Immunodeficiency. True positive and true negative cases were confirmed by experienced immunologists. All potential predictors available were used to derive ML models and this included 117 variables to train them. Three different ML models were compared with each other and to logistic regression analysis.

Participants were followed from the baseline visit 1 to the last visit for an average of 0.67 years with a mean of 2.2 visits during follow-up. Comparisons between the control group and the IEI risk group were examined. In the IEI risk subgroup, recurrent infections, autoimmune disorders, and suspicious laboratory findings were significantly more common reasons for referral to pediatric immunology (P = .004). The laboratory findings that were significantly more common in the IEI risk subgroup were anemia (P = .029), leukopenia (P = .028), neutropenia (P < .001), lymphopenia (P = .004), low IgA (P < .001), low IgG (P = .003) and increased levels of IgE (P = .02). According to the best ML model, the major variables associated with the primary outcome were IgA levels, scoring on the combined risk score, lymphopenia, age at first appointment, and altered serial complete blood count. ML models performed better than logistic regressions and showed high sensitivity and low specificity, suggesting a better performance for screening rather than diagnosing IEI.

The enhanced predictive power provided by ML models could be a resource to provide faster referral to immunology specialists, thereby avoiding delays in diagnosis and leading to better health outcomes.

The use of artificial intelligence is certainly going to be coming to Allergy and Immunology as it will with other medical specialties. This article is one of the first to demonstrate its utility in helping to screen for inborn errors of immunity. Although the mechanisms by which machine learning takes place may be somewhat obtuse for the practicing physician, it will be incumbent on us to recognize how AI can partner with us to improve patient care.