Common Variable Immunodeficiency Disorder: A Decade of Insights from a Cohort of 150 Patients in India and the Use of Machine Learning Algorithms to Predict Severity

Comprehensive Summary

Common Variable Immunodeficiency (CVID) is a disorder causing various immune infections due to low antibody production. The main objectives of this study were to develop an accessible AI model for predicting the disease’s severity and to identify common symptoms of the disorder. The sample set was composed of 150 CVID-diagnosed patients with an average age of 18 at a single care center in India, with most having a severe CVID presentation, and the most common symptom being multiple respiratory tract infections. Those with secondary hypogammaglobulinemia were excluded. According to Ameratunga’s severity score, which grades the severity of CVID complications on different organ systems, patients were placed into “severe” and “non-severe” groups. The median value of the patients’ cumulative scores was the threshold dividing the 2 groups. 64.6% were deemed severe and 35.3% as non-severe. All subjects completed a blood count panel, from which the Th/Tc ratio, CD4, CD16, class-switched memory B cells as well as serum IgG, IgA, IgM, and CD19 levels were utilized by 5 AI models (logistic regression, support vector machine (SVM), random forest (RF), lasso regression, and XGBost) to predict CVID severity. The most common clinical features in the cohort were recurrent infections (84.2%) which included respiratory infections (most common), infective diarrhea, and skin infections as well as an increased rate of cancerous growths. Regarding B-cell counts, the most common abnormality was a decrease in class-switched memory B cells, affecting 64.4% of patients, with 21.7% having a high deficiency. Gene sequencing identified monogenic causes in most tested patients, with LRBA deficiency being the most common for those tested in both pediatric and adult groups. Out of the 5 tested models, RF was superior (accuracy of 0.853, F1 score of 0.872) followed closely by XG Boost (accuracy of 0.824, F1 score of 0.833). All models ranked the Th:Tc ratio and CD19 as important predictive variables, while IgA was least important. In the future, the authors call for in-depth, large-scale genetic studies in the Indian population due to high rates of consanguineous marriages. They also advise increased awareness and research into the disease due to the significant number of adult CVID diagnoses in this study (51.3%) potentially highlighting delays in diagnosis. Lastly, they note that the AI models demonstrated robust predictive power and their identification of important predictive features may guide future data collection and classification tasks. The study was limited by the specificity of the dataset used.

Outcomes and Implications

The variable presentation of the disease, ranging from respiratory and GI infections to increased risk of cancer, increases diagnostic time and hinders prediction of disease severity. Many tests have been created to categorize patient symptoms and “score” disease progression, but cannot be easily used in clinical settings due to this diversity. This study aims to classify these symptoms to aid diagnostic ability especially for those in India, for which a study has not been done before. Diagnostic delays, shown from the high percentage of adult CVID diagnoses, corroborate this need. Prediction of CVID progression by an AI model may allow clinicians to identify high-risk patients and develop individualized treatment plans earlier but further testing is needed before implementation into the clinical setting.