Multidisciplinary genomic evaluation reveals adult inborn errors of immunity with rheumatic features

Inborn errors of immunity (IEI), formerly termed primary immunodeficiency diseases (PIDs), were initially recognized as disorders predominantly affecting children, characterized by severe and recurrent infections due to underlying genetic defects in immune function (1, 2). However, over the past decade, this concept has evolved substantially (3). IEIs are now recognized as a heterogeneous group of disorders encompassing not only immunodeficiency (ID) but also immune dysregulation syndromes, including autoinflammation, autoimmunity, severe allergy, and susceptibility to malignancy (4, 5). To date, over 500 causative genes have been identified, and novel gene–disease associations continue to be reported annually, underscoring the diagnostic and therapeutic challenges associated with these disorders.

With the expansion of the IEI spectrum, it is now well established that the diagnoses of these disorders are not confined to childhood. Recent epidemiological estimates indicate that over 50% of IEI cases may first present their characteristics after the age of 25, highlighting the clinical importance of adult-onset IEI (6, 7, 8). Moreover, the increasing recognition of late-onset entities—most notably VEXAS (vacuoles, E1 enzyme, X-linked, autoinflammatory, and somatic) syndrome—demonstrated that IEI can emerge de novo even in older individuals, particularly through acquired somatic variants of UBA1 (9). Compared to pediatric-onset IEI, adult-onset cases often exhibit atypical and gradually accumulating manifestations, along with diverse disease courses, resulting in delayed or missed diagnoses (7, 8, 10, 11).

Importantly, a subset of adult patients with IEI presents primarily with systemic inflammation or autoimmune manifestations—symptoms that mimic rheumatic and musculoskeletal diseases (RMD) (12, 13, 14). Indeed, emerging evidence indicates that these patients initially diagnosed with RMD may carry unrecognized monogenic immune defects (15, 16). This diagnostic ambiguity poses significant clinical risks; for instance, a delayed or incorrect diagnosis of RMD may lead to unwarranted use of immunosuppressive therapies, thereby predisposing patients with unrecognized ID to serious infections or further immune compromise (8, 17, 18, 19, 20, 21, 22, 23, 24, 25). Adding to this complexity, variants in autoinflammatory disease-related genes such as MEFV have been reported to modify RMD phenotypes, further underscoring that genetic findings must be interpreted with caution in clinical practice (26, 27, 28, 29). Therefore, physicians, particularly rheumatologists caring for adult individuals, must develop a better understanding of these conditions and collaborate across specialties to integrate genetic findings with clinical phenotypes, enabling timely and accurate diagnoses and management.

Genetic testing now plays a central role in the diagnosis of IEI. Advances in next-generation sequencing enable high-throughput, sensitive analysis of immune-related genes, improving diagnostic yield and facilitating mechanism-based decisions (30, 31). However, the utility of genetic testing in adults with RMD remains poorly characterized. The prevalence of pathogenic IEI-related variants in this population, and their implication for clinical management, have not been systematically investigated. Moreover, no standardized criteria currently exist to guide genetic evaluation in adult rheumatology practice.

To address these clinical questions, we conducted a comprehensive evaluation of the diagnostic yield and clinical impact of multidisciplinary genomic evaluation—including its influence on therapeutic strategy—in adult patients with suspected IEI presenting with RMD. Additionally, we investigated the utility of a structured clinical classification framework and machine learning–based approaches in predicting patients most likely to benefit from genetic evaluation.

A total of 50 adult patients met the inclusion criteria and were classified into three clinical groups as shown in Fig. 1. The mean age was 36.5 ± 13.8 years, and 68.0% were female (n = 34). Based on pretest clinical features, 17 patients were classified as ID, 24 as autoinflammatory disorders (AID), and 9 as non-ID/AID. Patients in the non-ID/AID group were often already diagnosed with RMD such as systemic lupus erythematosus or dermatomyositis (Table S1).

Table 1 summarizes the baseline characteristics across the three groups. The mean age at the time of genetic testing was similar in the ID and AID groups (mean 35.2 ± 8.2 years and 35.0 ± 15.7 years, respectively) but higher in the non-ID/AID group (mean 42.8 ± 16.5 years). Median disease duration was the longest in the AID group (8.0 years, interquartile range [IQR] 3.0–17.0), followed by non-ID/AID (7.0 years, IQR 1.8–10) and ID (3.0 years, IQR 1.1–23.5). A family history of similar symptoms was most prevalent in the AID group (37.5%), and less common in the non-ID/AID (11.1%) and ID (0%) groups.

The indication for genetic testing differed by groups: 88.2% of ID cases underwent testing for differential diagnosis, whereas 83.3% of AID patients were tested based on specific clinical suspicion. In contrast, testing in the non-ID/AID group was hypothesis-driven in one-third of patients and exploratory in the remainder.

Clinical manifestations reflected group-specific patterns. Periodic fever was a defining feature of the AID group. Other frequent symptoms included arthralgia (62.5%), chest or abdominal pain (45.8%), and enterocolitis (33.3%). Hearing loss was reported in two patients (8.3%), exclusively in the AID group. The non-ID/AID group exhibited high rates of nonperiodic fever (88.9%) and skin rash (44.4%). Across the 17 patients in the ID group, recurrent infections were observed in 10 patients, among whom upper respiratory tract infections were observed in seven patients (Table S2). Opportunistic and fungal infections occurred only in some patients (n = 5 and n = 2, respectively). Laboratory findings confirmed absent autoantibodies in ID, low frequency in AID (8.3%), and increased prevalence in non-ID/AID (33.3%), reflecting overlap with autoimmune phenotypes.

Pre-genetic treatment patterns aligned with clinical classification. Immunoglobulin replacement was administered in 70.6% of ID patients, colchicine in 58.3% of AID patients, and systemic glucocorticoids in 77.8% of non-ID/AID patients, indicating an empirical approach to disease management in this diagnostically ambiguous group. In addition, detailed immunological phenotype data were evaluated for patients who received immunoglobulin replacement therapy (Table S3). Serum IgA and IgM levels at the time of genetic testing were frequently low, with IgA below the lower limit of detection in several cases and markedly reduced in more than half of the patients. IgM levels were also decreased in the majority of cases, and a subset of patients exhibited profound hypogammaglobulinemia with both IgA and IgM below detectable limits, consistent with significant impairment of humoral immunity.

A final genetic diagnosis consistent with IEI was identified in 15 of 50 patients (30.0%) based on genetic findings and integrated clinical interpretation (Table 2 and Table S4). The most frequently affected gene was NLRP3 (n = 4), which is associated with cryopyrin-associated periodic syndrome, with all cases confined to the AID group. Other notable findings included haploinsufficiency of CTLA4 (n = 2), NFKB1 deficiency (n = 1), NFKB2 deficiency (n = 1), and NOD2 (n = 1), predominantly in the ID group. Two patients with UBA1 variants—diagnosed with VEXAS syndrome—were initially classified as non-ID/AID and identified via in-house sequencing. Additional rare variants were found in PRF1, GATA2, ALPK1, and ICOS. Based on genetic findings and the International Union of Immunological Societies (IUIS) 2024 phenotypic classification, clinical categories were reclassified in four patients: all were initially classified as non-ID/AID. Following reclassification, patients carrying variants in UBA1 or NOD2 were reassigned to the autoinflammatory diseases category, while one patient with CTLA4 was reclassified to diseases of immune dysregulation. No cases required reclassification from ID-related and autoinflammatory disease categories or vice versa.

Diagnostic yield was highest in the non-ID/AID group (44.4%). This may reflect the presence of previously unrecognized IEI within this heterogeneous category.

Genetic testing led to therapeutic modifications in 12 of 50 patients (24.0%) overall. Among 15 patients with a final genetic diagnosis, treatment changes were implemented in eight patients (53.3%) following genetic testing. In contrast, 4 of 35 patients without a final genetic diagnosis also underwent treatment modification based on clinical reassessment and phenotype-driven management. Immunoglobulin replacement was newly initiated in five patients—four in the ID group and one in the AID group. Canakinumab, an IL-1β inhibitor, was introduced in two AID cases, and systemic corticosteroids were introduced in one. Immunosuppressive agents were initiated in four patients (AID, n = 1; non-ID/AID, n = 3), reflecting the heterogeneity of immune dysfunction. Based on integration of clinical context and timing, 3 of the 12 interventions (25.0%) were classified as genotype-driven treatment, primarily corresponding to patients in whom genetic findings directly informed disease-specific therapeutic decisions (Table 3). The remaining nine interventions (75.0%) were considered predominantly phenotype driven, including treatment in patients with common variable ID without an identified monogenic defect, Blau syndrome, or VEXAS syndrome.

Two patients died during follow-up due to disease-related complications (Table S5). These included one ID patient with GATA2 deficiency and one non-ID/AID patient with VEXAS syndrome (UBA1). Causes of death were primarily infection-related, despite appropriate prophylactic treatment and extensive antimicrobial treatment. Both patients carried highly pathogenic variants and exhibited complex phenotypes, including pulmonary alveolar proteinosis, myelodysplastic syndrome, and Sweet syndrome, underscoring the clinical severity of adult-onset IEI.

Patients were classified into two groups according to genetic findings: genetically diagnosed, defined as those with pathogenic or likely pathogenic variants consistent with IEI, and genetically undiagnosed, defined as those without such variants. Compared with genetically undiagnosed patients, genetically diagnosed patients more frequently had a family history of the same disease (26.7 vs. 0%, P = 0.005) and less frequently had chest or abdominal pain (0 vs. 40.0%, P = 0.004) (Table 4). Chest or abdominal pain was less frequent (0 vs. 40%, P = 0.004) and lymphocyte counts were lower (mean 1,104/μl vs. 1,566/μl, P = 0.047) in patients with a final genetic diagnosis, while other clinical features were comparable between the groups. Colchicine use was more frequent in patients without a final genetic diagnosis (0 vs. 45.7%, P = 0.002), whereas other treatments did not differ. Moreover, no significant differences were found across ID, AID, and non-ID/AID classification with respect to these clinical variables.

Because conventional statistical analyses provided limited insight into the clinical backgrounds associated with a final genetic diagnosis, we further applied machine learning approaches as exploratory analyses to investigate clinical feature patterns and potential nonlinear relationships with genetic diagnosis in adult patients with RMD and suspected IEI. We first trained three gradient boosting classifiers (LightGBM, XGBoost, and CatBoost) on demographic, laboratory, and clinical features. Model performance was assessed by receiver operating characteristic–area under the curve (ROC-AUC), sensitivity, specificity, and accuracy. Among the three models, CatBoost showed the highest mean ROC-AUC performance, although variability across cross-validation folds was substantial (0.638 ± 0.250), as shown in Fig. S1. We then applied SHapley Additive exPlanations (SHAP) analysis to interpret the CatBoost classifier. SHAP summary plots revealed that chest or abdominal pain, family history of the same disease, autoantibody positivity, and lymphocyte counts were among the features contributing to model outputs (Fig. 2). SHAP dependence analysis further demonstrated that the presence of chest or abdominal pain was associated with lower SHAP values, indicating a reduced contribution to model outputs predicting a final genetic diagnosis (Fig. S2). For family history, higher scores were associated with increasing positive SHAP values, showing consistent directional effects across subgroups, although the magnitude of contribution tended to be greater among patients pre-classified as AID (Fig. S2). In contrast, autoantibody positivity demonstrated heterogeneous SHAP contributions across groups without a consistent directional shift, suggesting that its clinical relevance may depend on interactions with other clinical features rather than functioning as a strong discriminator (Fig. S2).

To visualize phenotypic overlap across the cohort, multiple correspondence analysis (MCA) was performed using dichotomized categorical variables. The MCA plot demonstrated considerable overlap between patients with and without a final genetic diagnosis, suggesting the difficulty of distinguishing these populations clinically. However, patients with a confirmed final genetic diagnosis showed relative proximity to family history (Fig. S3). Notably, genetically diagnosed patients were also located near cases in which treatment modification was implemented based on the diagnostic findings, whereas those without a genetic diagnosis were positioned closer to cases in which no treatment modification was made. This suggests that genetic diagnosis was closely associated not only with clinical stratification but also with subsequent changes in clinical management. To provide a clinically interpretable framework, we constructed a decision tree (classification and regression tree [CART] model) based on the same feature set. The resulting flowchart suggested a stepwise triage stratification structure with the initial split determined by lymphocyte counts and periodic fever, followed by branching based on chest or abdominal pain (Fig. S4). This rule-based approach illustrates how machine learning–derived feature patterns may support clinical stratification in diagnostically challenging adult patients.

This study investigated the clinical utility of multidisciplinary genomic evaluation in adult patients with suspected IEI who presented with RMD features—a clinical context in which the utility of genetic testing has not been well established. By focusing on these subsets, the study provides novel insights into the clinical application of genomic diagnostics in adult-onset IEI.

Given the considerable heterogeneity of clinical presentations and disease courses, IEIs are challenging to assess at the cohort level. Stratifying patients based on phenotypic presentation has been shown to improve diagnostic accuracy in previous studies (32).

In this study, we classified patients into three subgroups—ID, AID, and an atypical phenotype group (non-ID/AID). The ID group commonly lacked overt physical abnormalities but frequently presented with recurrent or opportunistic infections. The AID group exhibited features consistent with systemic autoinflammatory syndromes, such as arthralgia, serositis, and sensorineural hearing loss (33, 34).

The non-ID/AID group posed a notable diagnostic challenge due to its heterogeneous presentations, which often mimicked RMD without fulfilling the criteria for ID or AID. Many exhibited symptoms such as nonperiodic fever, arthralgia, and skin rash, and some tested positive for autoantibodies, complicating the distinction from classic RMD. Nevertheless, integrated genomic evaluation identified several adult-onset IEI in this group, including VEXAS syndrome (UBA1 variants) and CTLA4 haploinsufficiency, leading to disease reclassification according to the 2024 IUIS phenotypic framework (9, 35). These findings suggest that phenotype-based clinical classification alone may not fully reflect the underlying molecular pathology in adult-onset IEI. Importantly, several patients in the non-ID/AID group had already received empirical immunosuppressive therapy before genetic testing, underscoring the limitation of phenotype-driven management strategies in diagnostically complex adult patients.

Using pathogenic or likely pathogenic variants interpreted in the context of clinical phenotypes, we identified a final genetic diagnosis consistent with IEI in 30.0% of all patients, which was comparable to the diagnostic yield reported in previous adult IEI cohorts (25–30%) (31, 36). Although the number of genetically diagnosed patients within the non-ID/AID group was small, the relatively high diagnostic yield in this group (44.4%) may reflect the selective enrichment of diagnostically challenging adult patients in whom routine clinical assessment failed to establish a conclusive diagnosis. These patients subsequently underwent multidisciplinary evaluation and advanced genetic testing tailored to the clinical context, including in-house targeted sequencing, the Primary Immunodeficiency Database in Japan (PIDJ)-400 expanded panels, and whole-exome sequencing (WES) (7, 12, 37). For WES performed through the Initiative on Rare and Undiagnosed Diseases (IRUD) framework, variant interpretation undergoes additional review by an external expert panel composed of clinical geneticists and genome specialists who are independent from the primary clinical team, ensuring standardized American College of Medical Genetics and Genomics and the Association for Molecular Pathology (ACMG/AMP)–informed classification and improving diagnostic reliability (38, 39). In selected IRUD framework cases, trio-based analysis clarified inheritance patterns and contributed to variant interpretation, highlighting an additional benefit of family based evaluation even when relatives were not included as independent cases (40, 41). These approaches supported more reliable variant interpretation and clinical integration, suggesting that molecular diagnosis based on ACMG/AMP classification should be interpreted in the context of clinical phenotype, disease course, and treatment response.

In contrast, 35 patients (70.0%) remained without a final genetic diagnosis despite multidisciplinary evaluation integrating genetic findings with clinical phenotypes, possibly due to phenotypic variability, genetic heterogeneity, variants of uncertain significance or the presence of acquired phenocopies of IEI (42). The limitation of current sequencing technologies and incomplete gene coverage may also have played a role. Reanalysis of sequencing data, expansion of gene panels, and incorporation of functional studies may be required to elucidate etiology in these cases (7, 31, 36). Following a final genetic diagnosis, genotype-driven treatment modifications were identified in a subset of patients. In patients with ID, immunoglobulin replacement therapy was initiated following molecular confirmation, whereas in AID targeted therapies such as canakinumab were introduced based on underlying genetic findings. These findings highlight the potential clinical impact of genetic testing in guiding disease-specific therapeutic strategies. In contrast, for many adults with IEIs, including Blau syndrome (NOD2) and VEXAS syndrome (UBA1), therapeutic decisions remain primarily phenotype-driven because genotype-specific therapies are not yet established (43).

Despite diagnostic and therapeutic interventions, two patients died. Both had severe phenotypes associated with GATA2 deficiencies or UBA1 variants. In the previously reported GATA2 case, delayed recognition resulted in immunosuppressive treatment initiated as symptomatic management, subsequent severe and refractory opportunistic infections, ultimately precluding curative hematopoietic stem cell transplantation (20). These cases highlight that timely diagnosis is essential to prevent irreversible clinical deterioration (44, 45).

In conventional comparisons of patients with and without a final diagnosis, the presence of a family history of the same disease, chest or abdominal pain, and lymphocyte counts were associated with a final genetic diagnosis. To further explore potential nonlinear relationships among clinical features, we applied machine learning approaches. Because the cohort was preselected, these analyses are exploratory and highlight cohort-specific patterns rather than generalizable diagnostic indicators. The models visualized implicit clinical reasoning and identified chest or abdominal pain and family history as the features contributing to model outputs. These exploratory analyses should be interpreted cautiously because of the limited sample size and potential model instability. Taken together, they support the potential value of multidisciplinary evaluation in diagnostically complex adult cases and illustrate how genetic testing can meaningfully support diagnostic and therapeutic decision-making.

For patients who remain without a final diagnosis despite these efforts, continuous clinical follow-up and comprehensive immunological assessment should be considered to uncover hidden etiologies. Future studies should integrate additional biomarkers, such as cytokine profiles and molecular markers of lymphocyte neogenesis, including T cell receptor excision circles and κ-deleting recombination excision circles, and expand training datasets to improve model robustness and generalizability (46, 47, 48).

This study has several limitations. As a single-center retrospective analysis, potential selection bias influenced the distribution of IEI, and the diagnostic yield may be inflated because only patients undergoing genetic testing were included. Furthermore, the cohort was enriched for patients already considered highly suspicious for IEI based on prior clinical assessment and multidisciplinary evaluation, which may limit the generalizability of the findings. In addition, the inclusion of clinically recognizable classical autoinflammatory diseases—many of whom had already been clinically identified and treated prior to testing—may have led to an overestimation of the overall diagnostic yield. Because the diagnostic difficulty of classical autoinflammatory diseases differs substantially from that of other IEI, future studies should evaluate these groups separately. Moreover, in some cases, exact variant-level information was unavailable because of incomplete historical medical records during long-term follow-up. Finally, the relatively small sample size limited statistical power and may have contributed to model instability in exploratory machine learning analyses.

In conclusion, this study demonstrates that structured clinical classification combined with multidisciplinary genomic evaluation enables the identification of IEI in adult patients with RMD. Genetic testing not only facilitates precise diagnosis but also informs treatment in diagnostically challenging cases. These findings support consideration of genomic evaluation in adult rheumatology practices, especially in atypical or treatment-refractory disease presentations.

We conducted a retrospective observational study using data extracted from the clinical data warehouse of the Institute of Science Tokyo. Eligible patients were adults (≥18 years) who visited the Department of Rheumatology between March 2018 and September 2024 and were clinically suspected of having IEI. Clinical suspicion was based on the physician’s assessment, considering features such as recurrent infections, an atypical combination of symptoms, chronic inflammation, and/or a family history suggestive of immune dysregulation. Only patients who underwent genetic testing were included in the analysis. Incomplete data about genetic testing―defined as cases in which the performance of testing was unclear from the medical records, results were not documented in sufficient detail, or a final report was unavailable―were excluded from the study. The cohort included individuals referred from other adult departments, as well as patients who had previously undergone genetic testing in pediatric departments (either at our institution or externally) and were subsequently referred for evaluation by rheumatologists. To analyze patients with a final genetic diagnosis consistent with IEI, we excluded patients with familial Mediterranean fever carrying monoallelic variants of MEFV and those whose identified variants did not fulfill the ACMG/AMP criteria for pathogenic or likely pathogenic classification from downstream analysis (5, 49). A diagnostic flowchart summarizing the inclusion process and clinical classification framework is provided in Fig. 1.

A structured, multidisciplinary workflow was used to determine whether patients should proceed to genetic testing (Fig. S5). Patients with clinical features suggestive of classical monogenic autoinflammatory disease―such as MEFV- or NLRP3-associated disorders―were initially evaluated by rheumatologists, who ordered targeted genetic testing through the Kazusa DNA Research Institute (KDRI). When the clinical presentation was highly concordant with the suspected diagnosis, treatment decisions were made directly by the attending rheumatologist based on these results. However, if discrepancies between the clinical phenotype and the suspected diagnosis were noted, or if ID-like manifestations or recurrent infections were present, these classical autoinflammatory disease cases were also presented at a departmental multidisciplinary conference.

Patients with atypical inflammatory phenotypes, ID-like features, or symptoms insufficiently explained by conventional RMD were routinely reviewed at multidisciplinary conferences involving rheumatologists, pediatric immunologists, clinical geneticists, and genetic counselors. During these conferences, the likelihood of IEI, the potential therapeutic implications of the genetic findings, and the appropriate genetic testing modality were discussed.

Genetic analyses were performed using four platforms: (1) an in-house targeted panel, (2) targeted panels from KDRI, (3) the PIDJ-400 expanded panels, and (4) WES through the IRUD framework. Gene panels from KDRI were designed to cover well-established genes associated with ID and autoinflammatory diseases (Table S6). For broader coverage, we applied an expanded targeted sequencing panel covering ∼400 genes associated with IEI. This panel was developed under the PIDJ framework, based on the 2019 classification of IEI genes by the IUIS (50, 51). WES was performed within the IRUD framework, a nationwide diagnostic initiative in Japan. Trio-based sequencing comprising the proband and both biological parents is routinely implemented whenever feasible, enabling assessment of de novo variants, segregation, and inheritance patterns. As part of this framework, variant interpretation included an additional review by an external expert panel composed of clinical geneticists and genome specialists (38, 39). A complete flowchart of the testing workflow and selection logic is presented in Fig. S5.

Based on multidisciplinary evaluations, patients were directed to at least one of the available platforms: an in-house panel, KDRI panels, the PIDJ-400 expanded panel, or WES through IRUD. The PIDJ-400 panel and WES through IRUD were preferentially selected for cases in which broader differential diagnoses were considered or when initial targeted KDRI testing was inconclusive.

Variants were interpreted in accordance with guidelines of ACMG/AMP, using population allele frequency, Combined Annotation Dependent Depletion scores, and other supportive evidence (49). Variants classified as pathogenic or likely pathogenic according to ACMG/AMP guidelines were jointly reviewed by a multidisciplinary team comprising rheumatologists, pediatric immunologists, clinical geneticists, and genetic counselors. Final clinical interpretation was performed by integrating these molecular findings with phenotypes, disease course, and treatment response. In this study, all genetic diagnoses were subsequently categorized according to the 2024 IUIS phenotypic classification for human IEI (52).

Most of the patients in this study had several symptoms suggestive of RMD, such as musculoskeletal complaints, unexplained fever, or skin manifestations. Patients were classified into three groups based on their clinical presentation and treatment history:

• 1.

  ID: Patients with recurrent, persistent, or opportunistic infections and/or requiring regular immunoglobulin replacement therapy.

• 2.

  AID: Patients without clear features of ID but with periodic inflammatory episodes, including fever, chest or abdominal pain, skin rash, or sensorineural hearing loss.

• 3.

  Non-ID/AID: Patients who did not meet criteria for either ID or AID. This group included cases initially diagnosed with RMD such as undifferentiated connective tissue disease or seronegative arthritis. These patients exhibited atypical or overlapping features (e.g., nonperiodic fever, nonspecific skin rash, or negative or atypical autoantibodies) and unusual disease progression or poor response to conventional immunosuppressive/antirheumatic treatments.

Univariate analyses were performed to compare patients with and without pathogenic or likely pathogenic variants consistent with IEI. Categorical variables were compared using the chi-square test or Fisher’s exact test, as appropriate, and continuous variables were compared using the Mann–Whitney U test. Clinical data were aggregated and summarized using SPSS version 30.

We developed exploratory machine learning models to investigate clinical patterns associated with a genetic diagnosis consistent with IEI from our cohort data. Input variables consisted of demographic data, laboratory values (e.g., serum IgG levels and lymphocyte count), and clinical features (e.g., skin rash, autoantibody positivity, and family history). Prior to model training, feature engineering was performed (Table S7). For patients who were already receiving immunoglobulin supplementation therapy, serum IgG levels were coerced to 300 mg/dl. Missing IgG values were also imputed as 300 mg/dl. This threshold was chosen to approximate clinically relevant hypogammaglobulinemia as described in previous studies linking IgG < 400 mg/dl to PID (19, 53, 54). Family history was categorized into four ordered classes (same and similar, same only, similar only, or none), and history of infections was similarly categorized (frequent and recurrent, frequent only, recurrent only, or none). These ordinal categories were converted to integers to preserve ordinal relationships. Missing values were otherwise imputed using the median (for continuous variables) or the mode (for categorical variables).

Three gradient boosting implementations—LightGBM (v4.6.0), XGBoost (v3.2.0), and CatBoost (v1.2.8)—were trained on the dataset (55, 56). Model performance was evaluated using a nested cross-validation with a 15% independent holdout set. Fivefold outer cross-validation loops were used for model evaluation, with fivefold inner loops for hyperparameter optimization via Optuna (v4.7.0). Final models were trained on 90% of the data and tested on the holdout set. Discrimination was assessed by ROC-AUC as the primary metric, alongside sensitivity, specificity, and accuracy.

To enhance the interpretability of the machine learning model, we applied SHAP to the best-performing gradient boosting classifier (57). SHAP values were computed for each input variable to quantify its marginal contribution to the predicted probability of IEI. Summary plots were generated to visualize the overall impact, direction, and importance of each feature across the cohort.

To explore latent structures in categorical data, MCA was performed using features such as autoantibody positivity, skin rash, and family history (58). Patients with complete data (n = 50) were included. The results were projected in two dimensions to assess clustering patterns between genetically diagnosed and undiagnosed groups. MCA was performed in R (v4.5.2) using the MASS package (v7.3-65).

To develop a rule-based exploratory classification framework, we constructed a CART model using the same feature set (59). The Gini impurity criterion was used for node splitting, and tree depth was restricted to reduce overfitting. The resulting tree provided a stepwise, clinically interpretable algorithm for triage based on key variables.

All statistical and machine learning analyses were performed using Python (version 3.12.12) with XGBoost, LightGBM, CatBoost, SHAP, and prince libraries. The CART model was implemented in R (v4.5.2) using the rpart package (v4.1.24).

Supplementary information includes additional details regarding testing panels, preprocessing strategies, clinical characteristics, and supplementary analyses. Fig. S1 shows predictive performance of gradient boosting classifiers. Fig. S2 shows SHAP dependence analysis in relation to a final genetic diagnosis. Fig. S3 shows MCA of clinical features in adult patients with suspected IEI. Fig. S4 shows decision tree model for clinical stratification. Fig. S5 shows clinical workflow for multidisciplinary genomic evaluation in adults with suspected IEI. Table S1 shows presumed diagnoses assigned by primary physicians in the non-ID/AID group. Table S2 shows detailed infection history and prophylactic treatments in patients with the ID group. Table S3 shows serum immunoglobulin levels in patients receiving immunoglobulin replacement therapy at the time of genetic testing. Table S4 shows genetic variants identified in patients with a final genetic diagnosis. Table S5 shows a summary of fatal cases identified in the cohort. Table S6 shows diagnostic gene panels provided by KDRI. Table S7 shows variable preprocessing and encoding strategy for MCA and machine learning analyses.

This study was performed in line with the principles of the Declaration of Helsinki. This study was approved by the Institutional Review Board of the Institute of Science Tokyo (Approval No.: M2024-032). Informed consent was obtained using an opt-out approach, in accordance with institutional ethical guidelines for retrospective studies involving anonymized patient data. For patients undergoing genetic testing at the Institute of Science Tokyo, written informed consent was obtained individually. This included specific consent for the reporting of incidental genomic findings in accordance with ACMG guidelines. In addition, genetic testing conducted under the PIDJ-related research was separately approved by the Institutional Review Board of the Institute of Science Tokyo (Approval No.: G2019-004), and written informed consent was likewise obtained individually from each participant.

The novel disease-causing genetic variants identified in this study have been deposited in the ClinVar database under the following submission numbers: SCV007520436, SCV007520438, and SCV007612062.1. All data supporting the findings of this study are included within the article and its supplementary materials. Additional information is available from the corresponding author upon reasonable request. Individual-level WES data generated through the IRUD framework are not publicly available due to restrictions related to participant privacy, informed consent, and ethical approval for human genomic data.

The authors thank the clinical staff of the Department of Rheumatology, Institute of Science Tokyo, for their assistance in patient care and data acquisition. We further thank the Department of Pediatrics and Developmental Biology and the Department of Child Health and Development, Institute of Science Tokyo, for their collaboration and prior genetic assessments in selected patients. We also acknowledge the contributions of the Kazusa DNA Research Institute, the Primary Immunodeficiency Database in Japan consortium, and the Initiative on Rare and Undiagnosed Diseases in performing genetic analyses.

Author contributions: Hiroyuki Baba: conceptualization, data curation, formal analysis, investigation, methodology, project administration, resources, validation, visualization, and writing—original draft, review, and editing. Tadashi Hosoya: conceptualization, methodology, and writing—review and editing. Taiki Yamaguchi: conceptualization, formal analysis, methodology, software, visualization, and writing—original draft, review, and editing. Takuji Itakura: data curation and writing—review and editing. Hirokazu Sasaki: writing—review and editing. Natsuka Umezawa: investigation and writing—review and editing. Tetsuya Saito: conceptualization, data curation, project administration, and supervision. Naoki Kimura: resources and supervision. Ryuji Koike: investigation and writing—review and editing. Masayuki Yoshida: supervision. Masaki Shimizu: resources and writing—review and editing. Hirokazu Kanegane: supervision and writing—review and editing. Shinsuke Yasuda: funding acquisition, supervision, and writing—review and editing.