Enhancing the Prediction of Inborn Errors of Immunity: Integrating Jeffrey Modell Foundation Criteria with Clinical Variables Using Machine Learning

Background: Inborn errors of immunity (IEIs) are a heterogeneous group of rare disorders caused by genetic defects in one or more components of the immune system. The Jeffrey Modell Foundation’s (JMF) Ten Warning Signs are widely used for early detection; however, their diagnostic sensitivity is limited. Machine learning (ML) approaches may improve prediction accuracy by integrating additional clinical variables into decision-making frameworks. Methods: This retrospective study included 298 participants (98 IEI, 200 non-IEI) evaluated at a university-affiliated clinical immunology clinic between January and December 2020. IEI diagnoses were confirmed using European Society for Immunodeficiencies (ESID) criteria. Two datasets were constructed: one containing only JMF criteria and another combining JMF criteria with additional clinical variables. Four ML algorithms—random forest (RF), k-nearest neighbors (k-NN), support vector machine (SVM), and naive Bayes (NB)—were trained and optimized using nested 5-fold stratified cross-validation repeated three times. Performance metrics included accuracy, sensitivity, specificity, F1 score, Youden Index, and the area under the receiver operating characteristic curve (AUROC). SHapley Additive exPlanations (SHAP) were applied to evaluate feature importance. Results: Using only JMF criteria, the best-performing model was SVM (accuracy: 0.90 ± 0.04, sensitivity: 0.93 ± 0.05, AUROC: 0.91 ± 0.02). With the addition of clinical variables, the SVM achieved superior performance (accuracy: 0.94 ± 0.03, sensitivity: 0.97 ± 0.03, AUROC: 0.99 ± 0.00), outperforming both the classical JMF criteria (accuracy: 0.91, sensitivity: 0.87, AUROC: 0.90) and the JMF-only SVM model. SHAP analysis identified family history of early death, pneumonia history, and ICU admission as the most influential predictors. Conclusions: ML models, particularly SVM integrating JMF criteria with additional clinical variables, substantially improve IEI prediction compared with classical JMF criteria. Implementation of such models in clinical settings may facilitate earlier diagnosis and timely intervention, potentially reducing morbidity and healthcare burden in IEI patients.