A Point-of-Care Prediction Tool for Recurrent Tuberculosis

An estimated 10% of tuberculosis (TB) survivors who recently completed treatment in India develop TB again. We sought to develop a parsimonious model for predicting TB recurrence that can help target post-treatment active case finding among the highest-risk TB survivors.

The TB Aftermath trial enrolled TB survivors at treatment completion from 6 public TB clinics in Maharashtra, India, and assessed participants at 6-month intervals. Our prediction endpoint was recurrent TB diagnosed within 18 months of treatment completion. Candidate variables included risk factors for recurrence identified a priori and lung function assessments. We used LASSO (Least Absolute Shrinkage and Selection Operator) regression to shortlist predictors and estimated probability of recurrence using logistic regression. We conducted internal validation, assessed discrimination, and plotted calibration. Model selection was based on practical utility and predictive accuracy. For our selected model, we identified a cutoff for achieving 90% sensitivity.

Among 1033 participants, we identified 85 (8.2%) recurrences. Several 5-item models measurable at treatment completion had moderate discrimination. Our selected model included sex, household income, body mass index, peak expiratory flow from spirometry, and history of multiple TB episodes. The selected model had a cross-validated c-statistic of .69 (95% CI: .56-.77) and acceptable calibration (intercept: .03 [95% CI: -.03, .09]; slope: .66 [.08-1.24]). TB survivors with a predicted probability ≥3.7% accounted for 90% of recurrences.

A 5-item tool, measurable at treatment completion, showed moderate predictive accuracy for recurrent TB. At scale, a simple 5-item prediction tool may increase efficiency of post-treatment active case finding.