Records were identified by searching PubMed, Embase and Cochrane databases between January 2010 and November 2020, with the terms “Good syndrome”, “Good’s syndrome”, “thymoma” AND “hypogammaglobulinemia”, “thymoma” AND “immunodeficiency”, as well as “thymoma” AND “infection”. We chose 2010 as the initial year because a previous systematic review already summarized cases published between 1956 and 2009 (2). In addition, cases reported in the recent decade were described in a more standardized manner, allowing better synthesis of the data. We further reviewed the reference lists of retrieved records to identify additional cases. Duplicated records were excluded first. Two authors (YS and CW) screened the title and abstract of each record independently, and full texts of the record deemed relevant were reviewed by both authors to reach a consensus for inclusion or exclusion. As there is no formal diagnostic consensus for Good syndrome, we used the presence of both thymoma and hypogammaglobulinemia as the minimal criteria to define Good syndrome, consistent with the previous systematic review (2). Cases with individual patient data were included. Exclusion criteria were an inappropriate type of record (review, conference abstract or non-English publication), an alternative diagnosis or a study of Good syndrome with aggregated data. Aggregated data of the Good syndrome cohort were summarized and reported separately ( Supplementary Table 1 ) to allow comparison (3–6).

Primary aims of the current systematic review were to record the baseline clinical features (especially infection histories and related pathogens ( Supplementary Table 2 ), laboratory findings, concurrent diseases, treatments, and mortality of patients with Good syndrome. Given the diversity of pathogens reported in these patients, they were grouped according to the major type of host immunity defect predisposing to the infection, including humoral, cellular, phagocytic and others (including pathogens not associated with a particular type of immunity defects or pathogen was not reported for the infection) (7–11). Details of pathogens in each category were shown in Supplementary Table 3 . More advanced immunological workups, including vaccination response, lymphocyte proliferation to mitogens and respiratory burst test were not available in most cases. Prognostic factors impacting outcomes (mortality and survival) were explored. Clinical findings were also used to identify disease subgroups with similar manifestations and outcomes.

Data analysis was performed with SPSS version 26.0 (Armonk, NY, USA) and XLSTAT version 2020.1 version (New York, NY, USA). Continuous variables were summarized with median and interquartile range (IQR). Categorical variables were reported as numbers and percentages. Continuous variables were compared using Mann-Whitney test or Kruskal-Wallis test, and categorical variables were compared using Fisher’s exact test or chi-square test, as appropriate.

Binary logistic regression was used to identify prognostic factors impacting mortality for all patients (n = 162), and variables with both clinical and statistical relevance (p < 0.10) in univariate analysis were included in the multivariate model. For patients with clearly documented follow up duration (n = 109), overall survival was measured from the time of Good syndrome diagnosis until death or the last follow up. Survival curves were plotted using Kaplan-Meier method and compared with log-rank test. Cox proportional hazards model was further applied to test the significance of those prognostic factors regarding their survival impact.

To unravel homogeneous clinical subgroups within Good syndrome, a multiple correspondence analysis was first used to reduce the dimension of datasets. It cross-tabulated categorical variables and represented them graphically in a 2-dimensional Euclidean space by a multidimensional scaling technique (12). Input variables were chosen based on the defining criteria of Good syndrome, including thymoma, infection (type of pathogen and site of infection) and IgG level. After transformation of categorical variables into continuous variables (coordinates), we performed unsupervised hierarchical cluster analysis to determine clinical subgroups according to various characteristics. The clustering was constructed using Euclidean distance with the Ward agglomerative method. This method starts with each patient in its own cluster and the two most “similar” clusters based on Euclidean distance are combined at each step until the last two clusters are merged into a single cluster with all patients, as shown in a dendrogram (13, 14). Discriminative variables among clusters were further selected based on higher V test value with significant p value (15).

A decision tree was constructed with the use of chi-square automatic interaction detection technique to easily place the cases into different clusters (16). The p value was adjusted by Bonferroni correction. Ten-fold cross validation was applied to select the optimal tree. Performance of the decision tree was further evaluated by overall sensitivity and specificity of the cluster placement of cases.

All data were considered statistically significant at p < 0.05.

Our systematic literature review identified 162 patients from 121 records (17–137) ( Supplementary Figure 1 ). The demographics, clinical features and outcomes are shown in Table 1 . The median age of Good syndrome diagnosis was 58 years (IQR 51-67), and 51.0% were male. Because the ethnic origin of patient was not reported in most cases, geographic location of the corresponding author was recorded instead: Asia (n = 73, 45.1%), Europe (n = 58, 35.8%), North America (n = 22, 13.5%), Australia/Oceania (n = 5, 3.1%) and South America (n = 4, 2.5%). In 136 patients with available data regarding the chronological sequence of initial clinical presentation and subsequent Good syndrome diagnosis establishment, thymoma, infection and autoimmunity preceded the establishment of Good syndrome diagnosis in 44.9%, 29.4% and 25.7% patients, respectively. The median interval from initial clinical presentation to subsequent establishment of Good syndrome diagnosis was 24 months (IQR 4-72). Patients presented with infection (median interval 2 months, IQR 0-36) reached diagnosis earlier compared to those who presented with thymoma (median interval 36 months, IQR 6-82, p = 0.002) or autoimmunity (median interval 36 months, IQR 12-96, p = 0.001).

In 121 patients with thymoma histological classification data, WHO classification was used in 98 patients and type AB was found in 50.0%. Staging with the Masaoka system was only reported in 22 patients, and 68.2% were localized disease (stage I and II). At the time of Good syndrome diagnosis, half of the cases (51.4%) were in remission from thymoma standpoint.

Infections were recorded in 150 patients (92.6%), and the leading site was sinopulmonary tract (67.3%). Other common sites of infections included gastrointestinal system (27.3%), skin and soft tissue (18.0%), mucosa (14.0%) and eye (10.0%). Recovered pathogens are shown in Supplementary Table 2 . The most frequently recorded bacterium, fungus and virus were Pseudomonas spp. (12.7%), Candida spp. (16.7%), and cytomegalovirus (24.7%), respectively. Given the diversity of pathogens reported, they were further categorized according to the major type of host defect predisposing to the infection (7–11) ( Supplementary Table 3 ). As a result, 22.7%, 57.3% and 27.3% patients had infections related to humoral, cellular, and phagocytic defects.

Concurrent autoimmunity was identified in 83 patients (51.2%). Pure red cell aplasia (31.3%), myasthenia gravis (27.7%) and lichen planus (22.9%) were the most frequent disorders in these patients. Fourteen second pre-malignancy and malignancy were diagnosed in 13 patients (8.0%), before (n = 5), concurrent with (n = 3) and after (n = 6) diagnosis of Good syndrome. Eight were hematological and six were solid tumors.

All patients had hypogammaglobulinemia with a median IgG of 332 mg/dL (IQR 188-476) in those with reported levels (n = 126). Concurrent low IgA and IgM were noted in 86.0% (117/136) and 92.6% (125/135) patients with records. Absent or low peripheral B cells were shown in 118 of 124 patients (95.2%), as well as low absolute CD4 count in 77 of 108 patients (71.3%). Of 114 patients, 94 (82.5%) had an inverted CD4/8 ratio. The human immunodeficiency virus (HIV) was negative in all tested patients (n = 64).

There were no statistically significant differences in terms of immunoglobulin levels and peripheral B cell count between patients with and without infections related to humoral immunity defects. Regarding CD4 cell count and CD4/8 ratio, no significant differences were found between patients with and without infections related to cellular immunity defects ( Supplementary Table 4 ).

Of 150 patients with thymoma treatment data, 135 (90.0%) underwent thymectomy at the time of report. Chemotherapy and radiotherapy were given to 12 (8.0%) and 10 (6.7%) patients. In terms of immunoglobulin replacement, 127 patients (78.4%) received at least one dose. Antimicrobial prophylaxis was only used in 28 patients (17.3%). Of 83 patients with autoimmunity, 42 (50.6%) received concurrent immunosuppressive treatments (corticosteroid alone in 18 patients; immunosuppressants, such as cyclosporine, with or without corticosteroid in 24 patients).

A total of 25 patients (15.4%) died at the time of report. The main cause of mortality was infection (n = 23, 92.0%). In the whole cohort, factors impacting mortality consistently in both univariate and multivariate logistic regression included thymoma status (active disease, odds ratio [OR] 4.157, 95% confidence interval [CI] 1.219-14.177, p = 0.023), infection associated with cellular immunity defect (OR 3.324, 95% CI 1.100-10.046, p = 0.033), infection of sinopulmonary tract (OR 14.351, 95% CI 2.525-81.576, p = 0.003), infection of central nerve system (OR 6.403, 95% CI 1.205-34.027, p = 0.029) and infection of bloodstream (OR 6.917, 95% CI 1.519-31.505, p = 0.012) ( Table 2 ).

Of 109 patients with documented follow up duration (median 24 months, 95% CI 19-29 months), the 10-year overall survival was 53.7% (95% CI 25.8-75.1). The five prognostic factors identified by logistic regression in the whole cohort remained significant in these patients by both univariate and multivariate Cox regression ( Table 2 and Supplementary Table 5 ). Survival outcomes are depicted in Figure 1 .

Using multiple correspondence analysis and hierarchical clustering, we identified three clusters within the whole cohort ( Figure 2 and Supplementary Figure 2 ). Comparisons of clinical features among the three clusters are shown in Table 3 . Cluster 1 (n = 51, 31.5%) was best characterized by more frequent infections related to cellular immunity defect (94.1%, p < 0.001; V test, 7.071), affecting mucosa (37.3%, p < 0.001; V test, 6.220) and central nerve system (23.5%, p < 0.001; V test, 5.295). The most discriminative variables of cluster 2 (n = 77, 47.5%) included infections related to other immunity defect (61.0%, p < 0.001; V test, 4.033) and more frequent IgG < 500 mg/dL (88.9%, p = 0.041; V test, 2.269). In terms of cluster 3 (n = 34, 21.0%), patients had more frequent infections related to humoral (67.7%, p < 0.001; V test, 7.493) and phagocytic defects (58.8%, p < 0.001; V test, 5.041), as well as involving sinopulmonary tract (91.2%, p < 0.001; 3.891). The outcomes were statistically different between patients of cluster 1 and 2 (OR 0.325, 95% CI 0.118-0.893, p = 0.029; hazard ratio 0.382, 95% CI 0.150-0.974, p = 0.044 for those who with follow up duration) ( Supplementary Figure 3 ).

To easily classify cases into the three clusters, a decision tree was created by screening discriminative variables among the three clusters ( Figure 3 ). Using infections related to humoral (χ2 = 57.841, adjusted p < 0.001) and cellular defects (χ2 = 47.650, adjusted p < 0.001), the tree showed 72.2% correct estimation with an overall sensitivity of 72.3% and specificity of 86.2%.