Clinical Computer Vision for Tuberculosis Screening Under Portable X-Ray Device Heterogeneity: A Calibration-Aware Evaluation of Diagnostic Accuracy Across District Hospital Cohorts

Portable chest radiography has expanded tuberculosis screening in district hospitals, but the computer-vision classifiers deployed on these devices are typically trained and validated on images from stationary units, and their calibration under device heterogeneity remains uncharacterised. The framework distinguishes discrimination from calibration, formalises device-induced distribution shift as a mixture over acquisition strata, and derives a decomposition of pooled calibration error into device-mix and within-device components. The meta-analysis, estimated with the DerSimonian-Laird estimator, yielded a pooled effect of -2.19361 (SE 0.26513; 95% CI -2.71326 to -1.67395; z = -8.274; p = 2.22e-16) with I-squared of 99.2% and tau-squared of 1.45635, indicating that device-level heterogeneity of this magnitude would overwhelm any pooled calibration estimate that ignores stratification. The framework proposes that calibration reporting in tuberculosis screening should be stratified by acquisition device and accompanied by a device-mix sensitivity analysis, and it specifies the conditions under which pooled calibration claims are defensible.