Artificial Intelligence Tuberculosis Treatment Adherence Forecasting Under Variable Community Health Worker Smartphone Logging Gaps: A Missing-Not-At-Random Sensitivity Evaluation of Healthcare Loss-to-Follow-Up

Tuberculosis treatment adherence forecasting models trained on community health worker (CHW) smartphone logs face a structural problem: the logging process that generates the training signal is itself non-random, and the missingness mechanism is plausibly informative about the adherence outcome the model is meant to predict. The design parameters were fixed at between-cluster SD = 6.0, residual SD = 4.5, and intraclass correlation = 0.640; 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 is a design-characterisation study, not an empirical evaluation: no observational data are used or claimed. The framework distinguishes three logging-gap regimes (ignorable, covariate-dependent, and outcome-dependent) and derives the bias that each induces in the forecast target. The design analysis indicates that 56 clusters are required to detect a 5 percentage-point effect at 80% power under the declared variance structure, and that the minimum detectable effect at 80% power falls as cluster count rises from 8 to 32. The contribution is a sensitivity protocol that makes the MNAR assumption explicit and testable rather than silently absorbed into model training.