Abstract

Antibiotic-resistant tuberculosis (TB) strains pose a major challenge to TB eradication. Existing US epidemiological models have not fully incorporated the impact of antibiotic-resistance. To develop a more realistic model of US TB dynamics, we formulated a compartmental model integrating single- and multi-drug resistance. We fit twenty-seven parameters to twenty-two years of historical data using a genetic algorithm to minimize a non-differentiable error function. Since counts for several compartments are not available, many parameter combinations achieve very low error. We demonstrate that a crowd of near-best fits can provide compelling new evidence about the ranges of key parameters. While available data is sparse and insufficient to produce point estimates, our crowd of near-best fits computes remarkably consistent predictions about TB prevalence. We believe that our crowd-based approach is applicable to a common problem in mathematical biological research, namely situations where data are sparse and reliable point estimates cannot be directly obtained.