Temporal windowing of recurrent sinusitis improves EHR-based immunodeficiency classification

Recurrent sinopulmonary infections are a hallmark of common variable immunodeficiency (CVID), yet electronic health records (EHR) struggle to distinguish multiple discrete infections from a single infection comprising multiple clinic visits. Using diagnosis codes and antibiotic prescriptions, we developed a temporal windowing methodology that groups related sinusitis encounters into episodes based on timing and antibiotic escalation patterns. We applied this approach to 79 CVID patients, 224 rituximab-treated patients with hypogammaglobulinemia, and 2,994 matched controls. Immunodeficient patients had more frequent sinusitis episodes and greater episode burden than controls. Windowed episode features, when combined with infection PheCodes, yielded the best-performing Ridge regression classifiers in both the CVID-only (AUC-ROC 0.787) and pooled CVID/RTX (AUC-ROC 0.724) models, outperforming PheCodes combined with simple encounter counts and PheCodes alone. In a held-out cohort of five patients with pre-diagnosis sinusitis data, the classifier correctly flagged 4 of 5 as high-risk. Quantifying recurrence and refractoriness of infections through temporal windowing improves EHR-based phenotyping of immunodeficiency.