Development and internal validation of a clinical risk score for lost to follow-up among drug-resistant tuberculosis patients in a high-burden setting, Ethiopia: a multicenter retrospective study.

OBJECTIVES

Lost to follow-up (LTFU) in drug-resistant tuberculosis (DR-TB) treatment is a major global public health problem, leading to continued transmission, further amplification of drug resistance, and increased morbidity and mortality. Predicting the risk of LTFU is crucial for early intervention, but validated prediction models for DR-TB patients in Ethiopia are lacking. The aim was developing and internal validation of a risk prediction model for LTFU among DR-TB patients in selected treatment-initiating centers in the Amhara region, Ethiopia.

METHODS

A multicenter retrospective follow-up study was conducted using secondary data from 1097 DR-TB patients (2010-2025). Data were analyzed using R software. Predictors were selected using the least absolute shrinkage and selection operator (LASSO), and a multivariable logistic regression model was developed. Model performance was assessed by discrimination (area under the receiver operating characteristic curve-AUROC), calibration (Hosmer-Lemeshow test, calibration plot), and the Brier score. LASSO regression was performed with the penalty parameter (λ) optimized using bootstrapping-based internal validation. The optimal λ was determined by minimizing the mean prediction error across multiple bootstrapped resamples. Only internal validation was performed via bootstrap resampling, and clinical utility was evaluated with decision curve analysis.

RESULTS

The percentage of LTFU was 18.41%. The final prediction model included sex, nutritional status, residence, psychosocial support, counseling, distance to a health facility, major adverse drug events, and history of TB. The model showed excellent discrimination with an AUROC of 87.2% and good calibration (P = .239). Internal validation yielded an optimism-corrected AUROC of 86.8%. At an optimal probability cutoff of 0.27, the model had a sensitivity of 78% (95% confidence interval [CI]: 73-83) and specificity of 82% (95% CI :72%-86%). Decision curve analysis confirmed the model's clinical utility.

CONCLUSION

The percentage of LTFU was high. The developed nomogram is a robust and clinically useful tool for identifying DR-TB patients at high risk of LTFU.

PLAIN LANGUAGE SUMMARY: LTFU during DR-TB treatment can lead to poor outcomes and continued disease transmission. We developed and internally validated a prediction model using routinely collected clinical data to identify patients at high risk of LTFU. The model may help healthcare providers target early interventions to improve treatment completion. External validation is needed before widespread use.