Closing the U.S. Tuberculosis Care Gap: Artificial Intelligence, Video-Observed Therapy, and Specialist Workforce Alignment
TL;DR
Findings indicate that AI and VDOT can meaningfully extend diagnostic reach and treatment-monitoring capacity in a low-incidence U.S. setting, but that technological gains complement rather than substitute for the specialist workforce required to act on them.
Abstract
Tuberculosis (TB) remains a persistent, if low-incidence, public health problem in the United States, a low-incidence, high-resource epidemiological profile that differs fundamentally from the high-burden settings where most TB technology was originally developed. This paper synthesizes three linked lines of evidence bearing on how the U.S. TB care cascade can be strengthened: (1) a systematic review, structured around PICO criteria and PRISMA-style search and screening principles, of artificial intelligence (AI) and machine-learning (ML) applications in TB diagnosis and risk prediction; (2) a systematic review of real-world U.S. implementation evidence for video-observed therapy (VDOT) in treatment monitoring; and (3) a county-level ecological analysis linking California TB burden to infectious disease (ID) and pulmonology specialist workforce density. Across the AI literature, diagnostic computer-aided detection (CAD) software for chest radiography showed consistently high pooled sensitivity (approximately 85%–94%) but more variable specificity (approximately 37%–69%) across three independently published meta-analyses, while a published meta-analysis of VDOT found a significant pooled improvement in medication adherence relative to in-person directly observed therapy (DOT) (relative risk 2.79, 95% CI 2.26–3.45). Predictive AI applications, spanning progression-risk and outbreak-cluster models, were too heterogeneous across primary studies for quantitative pooling and are synthesized narratively. A deeper, complementary review of ten real-world U.S. VDOT implementation studies found that all seven observational studies reporting an adherence or completion outcome favored VDOT or found it comparable to DOT; a purpose-built meta-analysis combining the only two eligible U.S. active-disease studies with comparable completed-doses data (one observational, one randomized crossover) found a small but statistically significant risk difference favoring VDOT (+2.1 percentage points, 95% CI 0.1 to 4.1) and a larger but highly heterogeneous, non-significant difference for latent TB infection (LTBI) preventive therapy (+14.5 points, 95% CI −3.9 to 33.0, I²=74%), indicating that large gains in observation-rate metrics should not be assumed to translate proportionally into treatment completion. Only one of the ten studies reported a formal sociodemographic subgroup analysis, which found lower VDOT adherence among U.S.- or Mexico-born patients. Separately, an ecological analysis of 39 California counties found that population-adjusted TB case rate and ID/pulmonology physician density were weakly correlated overall (Pearson r=0.26) but moderately-to-strongly correlated once a single extreme outlier, Imperial County, was excluded (r=0.67); Imperial County combines the state's highest TB rate (25.0 per 100,000) with among its lowest specialist densities (0.56 per 100,000), a gap equivalent to roughly sixteen-fold relative to a comparable neighboring county. Because the specialist workforce gap identified here cannot be closed on any clinically meaningful timeline through routine training pipelines alone, these findings support prioritizing AI and VDOT adoption specifically in specialist-scarce jurisdictions: both technologies absorb high-volume, non-specialist components of the TB care cascade (screening triage, contact-cluster prioritization, routine dose observation), which can free scarce ID/pulmonology capacity for the complex, drug-resistant, and coinfected cases that require it, though this mechanism is inferred from the three analyses' individual findings rather than directly tested. Read together, these findings indicate that AI and VDOT can meaningfully extend diagnostic reach and treatment-monitoring capacity in a low-incidence U.S. setting, but that technological gains complement rather than substitute for the specialist workforce required to act on them, and that both technological and workforce interventions remain under-evaluated for the demographic and geographic subgroups who carry a disproportionate share of the nation's remaining TB burden.
