Global Health
Algorithmic Transportability of Pediatric Tuberculosis Treatment Models: The Impact of Measurement Heterogeneity and Missing Data Ieza Arshad* Ieza Arshad Arshad Arshad Boston University School of Public Health, Boston, USA
Background: Pediatric tuberculosis (TB) remains a diagnostic challenge due to nonspecific clinical presentations and limited bacteriologic testing. Clinical prediction models offer pragmatic tools for treatment decision-making, yet their transportability across diverse contexts remains uncharacterized. We assessed whether performance degradation in external settings is driven by patient characteristics and structural discrepancies such as predictor measurement heterogeneity and missing data mechanisms.
Methods: We validated two logistic regression pediatric TB models (derived in Pakistan and Bangladesh) using the INSPIRE dataset (a harmonized cohort of 4,718 children from 13 studies across 12 countries). We quantified the drop in discriminative performance (Youden’s J index) across settings. We evaluated the influence of inconsistent predictor definitions (e.g., “TB contact”) and conducted sensitivity analyses to identify the structural impact of missing data assumptions.
Results: Transportability was poor; Youden’s J declined by a mean of 0.42 in same-setting comparisons and averaged 0.27 in same-country cohorts. This was linked to data inconsistencies. For the Bangladesh model, specificity decreased by up to 67.5 percentage points (from 75.4% to 7.9%) in populations with different predictor distributions. Similarly, “pragmatic” missing data assumptions proved unstable: altering imputation from “missing=0” to “missing=1” reduced specificity by 17.4% in the Bangladesh model and 23.4% in the Pakistan model. Inconsistent definitions of “TB contact” and variable missingness (2.7% to 90%) introduced structural biases distorting risk scores.
Conclusion: Limited transportability stems from epidemiologic, measurement, and data-quality mismatches. Harmonized definitions, improved data completeness, and setting-specific recalibration are essential for equitable, reliable pediatric TB diagnostics. Future work should prioritize prospective validation and adaptive modeling.