Artificial Intelligence Benchmarking of Tuberculosis Drug-Resistance Prediction from Whole-Genome Sequences Under Unmeasured Mixed-Strain Heteroresistance: A Healthcare Calibration Evaluation

Whole-genome sequencing (WGS) is increasingly used to predict Mycobacterium tuberculosis drug resistance, and machine-learning classifiers trained on WGS data are benchmarked against phenotypic drug-susceptibility testing (DST). A persistent obstacle to fair benchmarking is mixed-strain heteroresistance: within a single patient, subpopulations of bacilli may carry different resistance mutations, so the phenotype recorded by DST is a population-level summary that does not correspond to any single genome. We define the calibration gap as the difference between the classifier's predicted probability of resistance and the probability that a randomly selected bacillus in the population is resistant, and we show that under a two-strain mixture model the gap is bounded by a function of the minor-variant allele frequency and the classifier's discriminative capacity. The framework is evaluated in a design study using a before-after-control-impact variance structure with between-cluster SD 6.0, residual SD 4.5 and intraclass correlation 0.64; power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis indicates that 56 clusters are required for 80% power at a 5-percentage-point effect. We argue that calibration, not discrimination alone, should be the primary benchmark for WGS-based resistance prediction, and we specify the reporting standards that would make heteroresistance-aware benchmarking reproducible.