Machine-Learning-Assisted Risk Stratification for Tuberculosis Infection in High-Risk Groups Receiving Immunosuppressive Therapy

Background: Individuals receiving immunosuppressive therapy, solid-organ transplant recipients, household contacts of patients with tuberculosis, and healthcare workers are at increased risk of Mycobacterium tuberculosis infection and progression to tuberculosis disease. However, currently available immunological tests cannot reliably distinguish tuberculosis infection from active tuberculosis disease or predict individual progression to disease. We aimed to assess the prevalence of immunological evidence of tuberculosis infection across clinically and epidemiologically distinct risk groups and to explore the potential of machine-learning methods for individualised risk stratification. Methods: In this observational study, 544 adults aged 18–75 years were assessed between 2023 and 2026. Participants included patients with pulmonary tuberculosis, tuberculosis-facility healthcare workers, healthcare workers in the general medical network, household and occupational contacts of patients with tuberculosis, patients with rheumatic diseases receiving immunosuppressive therapy, patients undergoing assessment for heart transplantation, heart-transplant recipients, and healthy controls. Participants underwent clinical, epidemiological, radiological, and immunological assessment, including interferon-γ release assays (ELISPOT and QuantiFERON-TB Gold) and the tuberculosis recombinant allergen skin test. Individuals with positive immunological tests underwent additional assessment to exclude active tuberculosis. Exploratory logistic regression and decision-tree models were developed separately for each risk group using demographic and treatment-related variables. Model performance was estimated using repeated stratified 5-fold cross-validation (5 × 10 repeats) with bootstrap optimism correction; discrimination, calibration, sensitivity, specificity, positive and negative predictive values, Cohen’s kappa, and Matthews correlation coefficient were reported with 95% confidence intervals. Results: The prevalence of positive immunological tests was highest among heart-transplant recipients (15 [53.6%] of 28), followed by household contacts (12 [41.4%] of 29), patients undergoing pre-transplant assessment (5 [35.7%] of 14), and tuberculosis-facility healthcare workers (21 [22.6%] of 93). Positivity was lower among patients with rheumatic diseases receiving immunosuppressive therapy (11 [13.9%] of 79) and healthcare workers in the general medical network without documented tuberculosis exposure (7 [10.5%] of 68), and was 3 (4.4%) of 69 among healthy controls. Compared with healthy controls, positivity was significantly higher among tuberculosis-facility healthcare workers, household contacts, patients with rheumatic diseases, and pre-transplant and heart-transplant populations. Exploratory machine-learning analyses were performed separately for the two clinically most relevant immunosuppressed subgroups. In patients with rheumatic diseases (n = 71; 9 positive, 62 negative), pooled ROC AUC was 0.670 (95% CI 0.520–0.824, participant-level bootstrap); age showed 100% bootstrap sign consistency, whereas bDMARD treatment showed 84.6% and sex 55.0%. In the combined heart transplantation group (n = 41; 19 positive, 22 negative), pooled ROC AUC was 0.668 (95% CI 0.522–0.811); tacrolimus and everolimus showed the highest bootstrap sign consistency (100% and 97.6%, respectively); however, because tacrolimus and everolimus are administered predominantly to transplant recipients, these coefficients may partly reflect transplantation status rather than drug-specific effects. Because the heart transplantation group combines pre-transplant candidates and recipients, and because both subgroups had small numbers of events, these associations should be interpreted as exploratory and not as drug-specific causal effects. Conclusions: The prevalence of immunological evidence of tuberculosis infection varied substantially between risk groups and was particularly high among household contacts and heart-transplant recipients. The findings suggest that epidemiological exposure and the type of immunosuppressive therapy should be considered jointly when assessing tuberculosis infection risk. Exploratory machine-learning analyses suggested that treatment-related variables may contribute to individualised risk stratification in the combined heart transplantation group, but the models were developed in small, clinically heterogeneous subgroups and were not externally validated. They should therefore be regarded as hypothesis-generating. Prospective multicentre studies with longitudinal follow-up are required to determine whether integrated clinical, epidemiological, immunological, and treatment-related models can predict incident tuberculosis disease and improve targeted preventive strategies.