Our study presents one of the largest cohorts of CVID patients in India, providing valuable insights into the clinical and immunological features of this heterogeneous disorder in the Indian context. While CVID is the most common symptomatic inborn error of immunity (IEI) globally, reports from India have been limited to isolated case reports or small series, often highlighting chronic diarrhea as a presenting symptom[1,2,3,4]. There are only two case reports with pulmonary manifestations in terms of recurrent pneumonia and bronchiectasis [5, 6].

The demographic profile of our cohort reveals both similarities and notable differences compared to previously reported large-scale studies from Europe and North America. The median age of diagnosis in our cohort was 18 years, with a wide range from 1 to 65 years. This broad spectrum highlights the heterogeneous nature of CVID and aligns with findings from other global cohorts, such as those reported by other large cohorts [7,8,9]. However, the significant proportion of adult diagnoses (51.3%) in our study highlights a potential delay in recognition and diagnosis of CVID in India, a critical issue that warrants increased awareness of this disease. The high rate of consanguineous marriages (22.2%) in our cohort is notably higher than in Western populations but aligns more closely with reports from Middle Eastern countries [10]. This finding underscores the potential role of genetic factors in CVID etiology within the Indian context and suggests a need for more comprehensive genetic studies in this population. The median delay in diagnosis of five years observed in our cohort which is comparable to other cohorts [9, 11].

Recurrent infections with involvement of the respiratory tract is the commonest clinical manifestation in our cohort as reported in most of the studies [8, 9, 12]. Most of the patients with lower respiratory tract infections had complicated pneumonias. 29 (19.3%) of these patients had developed bronchiectasis post infective sequelae. While several studies have reported lower IgM and IgA levels in CVID patients with bronchiectasis, our study did not find a statistically significant difference [13, 14]. This could be attributed to the high prevalence of severe pneumonia in our cohort, even among patients without bronchiectasis, potentially masking subtle differences in immunoglobulin levels. While Granulomatous–lymphocytic interstitial lung disease (GLILD) is a recognized complication of CVID, its detection in our cohort might have been limited by the non-specific presentation, and potential overlap with other lung diseases posing significant diagnostic challenges [15]. To improve GLILD detection, a high index of suspicion, early referral to specialized centres are crucial in these patients with overlapping features.

Gastrointestinal complications were frequent in our CVID cohort, affecting 45% of patients as reported in most of the cohorts [16, 17]. While infective diarrhoea was the most common presentation, the increased prevalence of chronic giardiasis (6.5%) and the presence of strongyloidiasis highlight the susceptibility to parasitic infections [18]. The occurrence of IBD and NLH further underscores the diverse GI manifestations in CVID and emphasize the need for a high index of suspicion, thorough diagnostic workup in such cases [19].

The significant proportion of patients (22.4%) in our CVID cohort required empiric anti-TB treatment, despite microbiological confirmation in only a small subset. The higher diagnostic delay in these patients (7 years) warrants a higher index of suspicion for CVID even in patients suspected to have complicated pulmonary TB. While early initiation of treatment is crucial, efforts should be made to obtain microbiological confirmation to minimize the risks of overtreatment.

The predominance of hematological autoimmunity, particularly Evans syndrome and Immune Thrombocytopenia (ITP), aligns with previous reports in 20–30% of CVID cases [20,21,22]. Lymphoproliferation, with or without hepatosplenomegaly, was observed in 33 patients, predominantly manifesting as cervical lymphadenopathy, which aligns with previous studies [9, 22]. The high incidence of NHL in CVID patients corroborates with previous studies [23]. This diverse range of malignancies underscores the increased cancer risk in CVID patients, and emphasizes the need for vigilant cancer screening and monitoring in this population [24,25,26].

Next-generation sequencing (NGS) in both pediatric and adult CVID patients revealed a high prevalence of monogenic causes, underscoring the genetic complexity of this disorder. Among the pediatric cohort, LRBA deficiency was the most common, followed by TRNT1 deficiency. In adults, while LRBA deficiency was also observed, CTLA4 and PIK3CD deficiencies were equally prevalent. These findings suggest potential age-related differences in the genetic landscape of CVID. However, it is important to acknowledge that some patients included in this study were later found to have molecular diagnoses of conditions that present with combined immunodeficiency disorders and diseases of immune dysregulation. These conditions, while sharing overlapping immunological and clinical features with CVID, are classified as CVID-like disorders after a definitive genetic diagnosis [10, 24]. This highlights the importance of distinguishing CVID from CVID-like disorders in genetic studies. Furthermore, the limited sample size of this study emphasizes the necessity for extensive NGS studies encompassing larger, diverse patient populations to fully elucidate the genetic etiology of CVID and potentially uncover novel genetic associations.

The high proportion of severe CVID cases in our cohort, as determined by Ameratunga's score, emphasizes the significant burden of this disease. The management approach in our cohort focused on individualized, symptom-based treatment, with co-trimoxazole prophylaxis and intermittent IVIG being mainstay therapies. However, cost constraints pose a significant barrier to optimal management, particularly for IVIG. The scenario will improve with inclusion of primary immunodeficiency disorders in the National rare disease policy 2021 under which patients will be getting IVIG replacement or HSCT free of cost [27]. The unfortunate deaths in our cohort underscore the potential severity of CVID, while the successful HSCT in one patient highlights the need for improved access to such potentially curative therapies in select cases. Further research is crucial to optimize management strategies and improve outcomes for CVID patients in India.

The immunophenotyping of peripheral blood subsets has given additional insights as has been described previously [28]. Several studies have evaluated the utility of different peripheral blood subsets and correlated the clinical presentation and progression [11, 28, 29]. The patients with very low B cell subsets and all except one of these had a severe phenotype. The defect in CSW B cells is the most commonly reported abnormality as has been reported previously highlighting the disturbed germinal center function in these patients [28,29,30,31]. A study has reported that patients with severe phenotype have higher IgG levels, CD8 T cells and lower CD4 T cells [32]. Lower IgA levels are associated with increased risk of bronchiectasis and other respiratory complications [33]. Higher IgM levels with lower B cell numbers is associated with reduced survival [34]. Low NK cells are reportedly seen in patients with severe bacterial infections and granulomas and have a higher mortality [35]. Specific antibody response is another important immune assay that is associated with the disease severity. Almost all patients referred to us are symptomatic at presentation requiring IVIG, assessing the response becomes difficult in them.

VISUAL score is a prognostic laboratory score based on, CD4 T cells, CSW B cells, serum IgA, IgG, and IgM levels at diagnosis [36]. While the VISUAL score offers a promising approach to predicting CVID severity by incorporating immunological parameters, its reliance on specific antibody responses poses a significant limitation. In many clinical settings, obtaining and interpreting these specific antibody responses can be challenging due to factors like resource availability, testing complexities, and the often-nuanced nature of antibody results [8, 9]. This limitation can hinder the widespread applicability and practical implementation of the VISUAL score in routine clinical practice.

In our study we utilized the serum IgG, IgA, IgM, CD19, CSW B cells, CD4, Th:Tc ratio, and CD16 cell counts to predict the severity based on the Ameratunga severity score. We compared the performance of five machine learning models in predicting CVID disease severity. Our findings robustly demonstrate that ensemble methods, particularly Random Forest and XGBoost, outperform other models across multiple performance metrics in this specific context.

The superior performance of Random Forest and XGBoost can be attributed to their ability to capture complex, non-linear relationships between features and their inherent resistance to overfitting. These characteristics likely contributed to their enhanced predictive power, especially given the complexity and potential noise in our dataset.

While simpler models like logistic regression and Lasso showed moderate performance, their computational efficiency and interpretability may make them valuable in scenarios where these factors are prioritized over raw predictive power. The SVM model's higher F1 score despite lower accuracy highlights the importance of considering multiple performance metrics, particularly in imbalanced dataset scenarios.

The Th:Tc ratio consistently ranks as a highly important variable across all models, suggesting its strong predictive value. Similarly, CD19 consistently demonstrates notable importance. Conversely, IgA consistently ranks as least important. While other variables like IgM, CSW, CD16, IgG, and CD4 fluctuate in importance across the models, they highlight the potential value of a multi-parameter approach in predicting CVID severity.

The identification of key predictive features through the Random Forest model's feature importance analysis provides valuable insights that could guide future data collection efforts or feature engineering in similar classification tasks.

Limitations of this study include the specific nature of the dataset used, which may limit the generalizability of our findings to other domains or datasets with different characteristics.

In conclusion, we report a large cohort of CVID patients diagnosed over a decade from India. The patients present with a diverse clinical phenotype and mostly are diagnosed at a severe stage. This study provides a robust comparison of machine learning models for binary classification, offering valuable insights to predict the severity in CVID patients using basic immunophenotypic workup.