Large-scale deep learning analysis for the early diagnosis of primary immunodeficiencies

Abstract Primary immunodeficiency (PID) is a group of heterogeneous disorders resulting from immune system defects. The early PID diagnosis is compromised by the heterogeneous manifestations along with low clinical awareness. Most PID cases are significantly underdiagnosed leading to increased mortality, co-morbidities and healthcare visits and costs. Among PID, combined immunodeficiencies (CID) are characterized by complex immune defects. Common variable immunodeficiency (CVID) is among the most common types of PID. In light of available treatments for CID and CVID, it is critical to systematize their early diagnosis. Our study objectives were two-fold. First, we developed and evaluated an accurate deep learning model to analyze administrative medical claims data from EHRs towards systematizing screening and early identification of CID and CVID. Second, we revealed the most important CID- and CVID-associated clinical phenotypes and their combinations, demonstrating a systematic methodology to improve early identification of these PID. All data were composed of medical claims derived from the Optum® de-identified electronic health record (EHR) database. Four large cohorts were generated: 797, 797, 2,312, and 19,924 CID/CVID cases and equal control sizes in Cohorts 1–4, respectively (a total of 47,660 cases and controls). Two deep learning models were developed (TabMLPNet and TabResNet) and compared against baseline models. Univariate logistic regression was used to calculate odds ratios across all clinical phenotypes and their combinations. The TabMLPNet model showed the highest diagnostic performance across cohorts with sensitivity, specificity, and overall accuracy ranging from 0.82–0.88, 0.82–0.85, and 0.80–0.87, respectively. For the first time, we identified distinctive combinations of antecedent phenotypes associated with CID/CVID per cohort, being consisted of respiratory infections/conditions, genetic anomalies, cardiac defects, autoimmune diseases, blood disorders and malignancies. Most phenotypes emerged were well described in the literature, which validated our findings. Moreover, several less well documented individual phenotypes (i.e., asthma, coagulation defects complicating pregnancy, cancer of lymphoid histiocytic tissue, lymphoid leukemia chronic) were also identified, which can lead to better clinical surveillance of PID. We demonstrated a generalized and accurate method evaluated on a large EHR-derived cohort of CID/CVID cases and controls. Our methodology can lead to the development of new clinical guidelines and pathways for earlier identification of the most important antecedent phenotypes and their combinations, enhance clinical awareness and be used to improve PID diagnosis and outcomes on a population level.