Computational validation of a CDS tool for LAI‑PrEP bridge‑period navigation

A modeling paper reports computational validation of a configurable clinical decision support (CDS) tool described as a LAI‑PrEP bridge‑period CDS algorithm intended to reduce attrition between long-acting injectable PrEP prescription and first injection, otherwise known as the "bridge period."

At the Tier‑4 synthetic population scale aligned to UNAIDS targets, the authors report a baseline bridge-period success rate of 23.96% (95% CI, 23.94–23.98%), rising to 43.50% (95% CI, 43.48–43.52%) under evidence-based intervention bundles. They describe this as an absolute improvement of 19.54 percentage points, corresponding to about 4.1 million additional successful transitions from prescription to first injection at target scale. These figures are presented as model outputs generated under the study’s specified assumptions at UNAIDS-scale volumes.

The authors describe the CDS as an externally configurable decision support tool that synthesizes evidence from major LAI‑PrEP trials (including HPTN 083, HPTN 084, and the PURPOSE program) and implementation studies to support risk stratification, barrier identification, and intervention selection. A core design element is a library of 21 interventions paired with mechanism-diversity scoring, which the paper says is intended to reduce redundant recommendations when multiple options target similar causal pathways. Validation is described as “progressive,” moving through four synthetic population tiers (1K, 1M, 10M, and 21.2M) to assess convergence and precision as scale increases. As a computational check, the paper reports comprehensive unit testing across 18 edge cases with a 100% pass rate, and it distinguishes this from prospective clinical validation in real-world care settings. In this framing, computational validity is defined by mathematical correctness, stability across scales, and internally consistent outputs under fixed inputs.

Using the translation approach outlined in the paper and parameters the authors treat as epidemiologic assumptions—HIV incidence of 2–5% among indicated users and LAI‑PrEP efficacy of 96%—the model converts the added successful transitions into downstream population projections. These figures are conditional model outputs based on the stated assumptions, not observed outcomes from clinical implementation. Under these assumed inputs, the authors estimate approximately 80,000–100,000 HIV infections prevented annually (midpoint: 90,000), alongside an estimated USD 40 billion in averted lifetime treatment costs. They situate the 21.2 million-person scenario as corresponding to what the paper describes as the UNAIDS 2025 PrEP target of 21.2 million and present the infections- and cost-related figures as conditional on the incidence and efficacy parameters used in the model. The paper emphasizes that these projected impacts are derived outputs of the modeled pipeline rather than empiric observations.

At the same target scale, the paper reports substantial differences in baseline bridge-period success by population and region. Across key populations, baseline success is reported to range from 10.36% for people who inject drugs (PWID) to 33.11% for men who have sex with men (MSM), a 22.75 percentage-point gap. By region, baseline success is reported to range from 21.69% in Sub-Saharan Africa to 29.33% in Europe/Central Asia, a 7.64 percentage-point gap. Against that baseline landscape, the authors report that intervention bundles yielded the largest relative improvements among groups starting from the lowest baseline success, highlighting PWID (+265% relative improvement) and adolescents (+147% relative improvement) as examples. This pattern is presented as a modeled distribution of gains, concentrating relative change where baseline bridge-period completion is lowest.

The paper also reports a barrier-focused analysis in which bridge-period success declines as structural barriers accumulate, described as a dose–response pattern with a threshold effect at higher barrier counts. In the reported summary statistics, success is 44.0% with zero barriers; each additional barrier reduces success by an average of 7.74 percentage points (with diminishing returns at higher counts); and 85.7% of the Tier‑4 synthetic population (18.2 million people) is modeled as having at least one barrier. The authors also report that 3+ barriers correspond to <15% success without interventions.

Alongside these scenario findings, the manuscript notes that computational checks and scale consistency are the primary validation artifacts described, rather than individual-level clinical classification metrics such as sensitivity or specificity for predicting completion. The authors explicitly separate computational precision from prospective real-world validation, framing the former as necessary but not sufficient for establishing clinical validity.

Key Takeaways:

• In a Tier‑4 synthetic population model, the authors report bridge-period success increasing from 23.96% to 43.50% with evidence-based intervention bundles, an absolute change of 19.54 percentage points and about 4.1 million additional successful transitions at target scale.
• Using assumed HIV incidence (2–5%) and LAI‑PrEP efficacy (96%), the authors estimate roughly 80,000–100,000 annual infections prevented and about USD 40 billion in averted lifetime treatment costs.
• Reported baseline disparities by population (PWID vs MSM) and region (Sub-Saharan Africa vs Europe/Central Asia) coincide with modeled patterns of larger relative improvements among groups with lower starting success, while the authors emphasize computational validity does not substitute for prospective validation.