PrEP reinitiation, uptake, and persistence as targets for reducing HIV incidence and racial and ethnic inequities among men who have sex with men in Boston, USA: a network modeling study

Summary

Background

A decade of pre-exposure prophylaxis (PrEP) has lowered HIV incidence among men who have sex with men (MSM) but not enough to meet Ending the HIV Epidemic goals, and racial and ethnic disparities persist. The PrEP era has produced a population of former users facing lower barriers to restarting than PrEP-naive individuals. Reinitiation has not been treated as a distinct target. We aimed to compare its population-level impact with that of uptake and persistence.

Methods

We used a network model of 100,000 cisgender MSM aged 15–65 in Boston, USA (21.5% Black, 18.9% Hispanic, 44.5% White, and 15.1% Other), parameterized with electronic health record data from 37,722 MSM (2012–2023) and calibrated to local surveillance. We compared 10-year population-wide and equity-focused (Black and Hispanic MSM) interventions scaling uptake, persistence, and reinitiation, reporting infections averted, number needed to treat (NNT), and disparities at comparison intensities, defined as multiples of the baseline reinitiation rate.

Findings

A 2× reinitiation scale-up averted 15.8% of infections (1481; NNT 40) versus 10.0% for 2× uptake and 5.6% for maximum persistence; 5× reinitiation averted 32.6% at the same NNT. Equity-focused reinitiation produced the most favorable relative disparity patterns of the three levers, though inequities remained.

Interpretation

Reinitiation was the most impactful PrEP continuum lever, a prevention-side analog of ART re-engagement warranting prospective evaluation. Restart can be supported with existing infrastructure, including EHR-triggered prompts at return encounters, and equity-focused implementation for Black and Hispanic MSM offers the most favorable disparity patterns.

Funding

US National Institutes of Health and the Emory Center for AIDS Research.

Research in context.

Evidence before this study

We searched PubMed and Web of Science from January 1, 2010, to March 1, 2026, for systematic reviews and primary studies of PrEP discontinuation, PrEP reinitiation, and modeling studies of the PrEP care continuum, using combinations of the terms “pre-exposure prophylaxis,” “PrEP,” “reinitiation,” “restart,” “discontinuation,” “persistence,” “mathematical model,” and “agent-based model,” restricted to studies in men who have sex with men in the United States. Two systematic reviews characterized reinitiation as a measurable phenomenon (pooled reinitiation 24%–37% after discontinuation) but found zero intervention trials targeting reinitiation. Prior agent-based modeling studies of PrEP and HIV transmission have universally represented PrEP restart as a memoryless process, with prior users re-entering the eligible pool at the same rate as PrEP-naive individuals, conflating initiation and reinitiation as intervention targets.

Added value of this study

To our knowledge, this is the first calibrated mathematical model that separates PrEP initiation and reinitiation as distinct race- and stopper-class-specific rate processes parameterized from electronic health record data on the same population. PrEP reinitiation interventions averted substantially more HIV infections than uptake or persistence interventions at comparable intensities; subgroup analyses identified short initial PrEP episodes (under 27 weeks) as a clinically actionable, modestly higher-impact restart context than longer episodes. Population-wide implementations of any of the three levers tended to leave or widen relative racial and ethnic disparities, whereas equity-focused implementations reduced incidence in Black and Hispanic MSM and produced more favorable disparity patterns, with equity-focused reinitiation showing the most favorable relative disparity ranges among the three levers in the full sweep, although a clear monotonic dose–response was not observed.

Implications of all the available evidence

PrEP reinitiation is an under-recognized and disproportionately impactful lever on HIV incidence among MSM in mature PrEP environments. Routine clinical and public health infrastructure, including EHR-triggered prompts at return encounters and pharmacist-led PrEP services, can be repurposed to support reinitiation as a primary intervention target. Equity-focused reinitiation interventions delivered to Black and Hispanic MSM offer a distinct route to both incidence reduction and equity, but cannot close the gap on their own; structural drivers of disparity, including health system access, insurance continuity, and structural racism in care delivery, must be addressed in parallel. Prospective evaluation of reinitiation-specific intervention strategies is warranted.

Introduction

HIV pre-exposure prophylaxis (PrEP) is highly effective but remains inequitably implemented in the United States.^1,2 Black and Hispanic men who have sex with men (MSM) continue to bear a disproportionate share of new HIV infections despite more than a decade of PrEP availability; in Massachusetts, PrEP use has remained concentrated among White MSM.² The PrEP care continuum, as articulated by Nunn and colleagues, conceptualizes the user's journey as awareness, uptake, adherence, and persistence.³ Racial and ethnic disparities have been documented at every step of this continuum.^4,5 Most intervention efforts to date have focused on naive initiation, and on adherence or retention among current users.^6,7

Reinitiation, restarting PrEP after a prior discontinuation, is a fifth step of the PrEP care continuum and has been the least studied. Discontinuation is common: cohort studies estimate that 30%–60% of MSM discontinue PrEP within the first year.8, 9, 10 HIV incidence after discontinuation is up to 7.5 times higher than during active PrEP use.¹¹ A 2022 systematic review and meta-analysis of PrEP discontinuation identified pooled reinitiation proportions of 24%–37%, but treated reinitiation only as an observed downstream outcome of discontinuation, not as an intervention target.¹² A 2024 systematic review of reinitiation specifically identified only 30 studies worldwide, and zero trials of interventions targeting reinitiation.¹³ Prior PrEP users are substantially more willing to use PrEP than people who have never used PrEP.¹⁴ They typically retain a clinical relationship, prior labs, and insurance authorization, and report lower stigma and disclosure barriers than first-time initiators.¹⁵ Reinitiation is therefore best understood within the broader framework of retention and re-engagement, yet it is a distinct, separately measurable continuum step and a tractable intervention target in its own right.

PrEP programs in the United States are now more than a decade old, and clinical cohorts increasingly contain large numbers of prior users who are eligible to restart. Because the reinitiation-eligible population is bounded by those who have ever initiated PrEP, this lever has the greatest potential in mature settings with a large accumulated pool of former users, and limited reach where cumulative uptake remains low. Public health programs and clinics therefore face a recurring allocation decision: whether to invest scarce intervention resources in initial PrEP uptake campaigns, in retention and persistence support for current users, or in re-engagement of prior users. This decision is especially pressing for Black and Hispanic MSM, in whom uniformly applied continuum interventions risk widening rather than closing relative inequities. Mathematical modeling has been the primary quantitative tool for comparing HIV prevention strategies before trials,16, 17, 18, 19, 20, 21, 22 but prior models have represented PrEP restart as a memoryless process, with former users re-entering the eligible pool at the naive initiation rate. This representation conflates initiation and reinitiation, and obscures the empirical observation that restart rates among former users can exceed naive-user rates by an order of magnitude in mature PrEP environments. As a result, the population-level impact of reinitiation interventions is unknown.

In this study, we used a race/ethnicity-stratified network-based mathematical model of MSM in the Boston metropolitan area, parameterized with longitudinal electronic health record (EHR) data from three Boston-area health systems. The model separated initiation and reinitiation as distinct rate processes, so that interventions on each continuum step could be evaluated independently and contrasted head-to-head. We compared the potential 10-year impact of PrEP uptake, persistence, and reinitiation interventions on HIV incidence and on racial and ethnic inequities. For each continuum lever we simulated a population-wide (equal-implementation) scenario and an equity-focused scenario in which the intervention was tailored to Black and Hispanic MSM. To our knowledge, this is the first modeling study to treat PrEP reinitiation as a distinct implementation target, to parameterize initiation and reinitiation from the same population's EHR data.

Methods

Model overview and population

We used EpiModelHIV, an open-source stochastic network model of HIV and other sexually transmitted infections (STIs) among MSM.¹⁶ The model simulates 100,000 sexually active MSM aged 15–65 in the Boston metropolitan area, stratified by race/ethnicity (Black, Hispanic, White, and Other). The population is open: individuals enter at sexual debut (age 15) and exit through age- and race/ethnicity-specific mortality or at age 65, with entries and exits balanced to hold population size approximately stable. Partnership formation and dissolution are modeled using temporal exponential random graph models (TERGMs)²³ fit to egocentric network data from ARTnet.²⁴ Partnerships are stratified by type (main, casual, and one-time) and by race/ethnicity- and age-mixing from ARTnet. ARTnet is a US egocentric survey of MSM (2017–2019; 4904 respondents reporting on 16,198 partnerships) used to estimate partnership degree, mixing, and durations, with a geographic term generating Boston-specific predictions. Per-act transmission probabilities depend on serostatus, viral load, condom use, anatomic role, STI status, and PrEP status. HIV natural history includes stage-specific viral load trajectories that drive infectiousness, antiretroviral therapy initiation, adherence, and viral suppression dynamics. The model was implemented in R version 4.4.2 using the EpiModel (version 2.5.0) and EpiModelHIV (version 3.1.0) packages. The model was calibrated to Boston-specific HIV prevalence, HIV status awareness, linkage to care, viral suppression, STI incidence, and PrEP coverage targets using a two-stage, surrogate-assisted Bayesian procedure and 2018–2023 surveillance data.^2,25,26 Full specification of the network, natural history, and clinical parameters appears in Supplementary Sections S1–S13 (Supplementary Tables S1–S14; Supplementary Figures S1–S8). Simulated values reproduced the calibration targets (Supplementary Figures S4–S8; Supplementary Tables S10–S13).

Representation of the PrEP care continuum

PrEP use follows an initiation, persistence, discontinuation, and reinitiation cycle (Fig. 1). PrEP-eligible individuals initiate at a race-specific weekly rate estimated from Boston EHR data among PrEP-naive users. Eligibility used CDC clinical indication criteria applied weekly based on recent sexual behavior, partnership characteristics, and STI history. On PrEP, individuals were assigned one of three adherence classes (low, medium, and high) at initiation, with class-specific per-act efficacy modifying HIV transmission at each exposure act.¹⁷ Each agent was pre-assigned at cohort entry to one of two stopper classes—fast or slow—with race-specific probabilities estimated from EHR data. Fast and slow stoppers faced class- and race-specific constant weekly discontinuation hazards, calibrated so the fast-stopper class matched the EHR sub-population with time on PrEP under 27 weeks (two consecutive 3-month fills) and the slow-stopper class those at or above 27 weeks. Mean initial-episode durations were 12–14 weeks for fast stoppers and 67–82 weeks for slow stoppers across race/ethnicity groups (Table 1). This prescription-coverage measure is conceptually aligned with the proportion of days covered rather than with a daily versus on-demand regimen label. This is a clinically grounded prescribing boundary, not a statistical one: two consecutive 90-day daily-oral prescriptions are the standard US dispensing unit, so the cut-point cleanly separates rapid early drop-off from sustained early use, and it distinguishes two discontinuation patterns with distinct post-stop reinitiation rates, consistent with prior trajectory analyses of PrEP persistence cohorts.²⁷ After discontinuation, individuals enter a race- and stopper-class-specific reinitiation process, so restart rates depend on race/ethnicity and on whether the prior discontinuation was short- or long-term. Unlike prior models that treat restart as memoryless, it maintains each agent's discontinuation history. Reinitiated individuals remain subject to the same discontinuation and adherence processes, so repeated cycles arise endogenously. This enables direct parameterization of reinitiation from EHR data and lets initiation and reinitiation be manipulated independently as intervention targets.

Fig. 1.

Schematic of PrEP initiation, discontinuation, and reinitiation in the network model. At each time step, a PrEP-eligible MSM is routed to one of three states based on prior PrEP use history and a pre-assigned stopper class: never used PrEP (initiating at the race-specific initiation rate), prior fast-stopper class (short initial episode, restarting at the race-specific fast-stopper reinitiation rate), or prior slow-stopper class (longer initial episode, restarting at the race-specific slow-stopper reinitiation rate). Fast and slow stopper classes are fixed agent attributes drawn at cohort entry from race-specific EHR episode-duration patterns; the 27-week threshold shown in the schematic distinguishes the two observed classes rather than an in-simulation duration switch. Unlike prior agent-based models that treat restart as memoryless, this structure maintains an individual's prior-use history and parameterizes initiation and reinitiation as distinct processes. Individuals may undergo repeated discontinuation–reinitiation cycles over the simulation, remaining subject to the same class-specific hazards after each restart.

Table 1.

Race/ethnicity-specific PrEP care continuum parameters and intervention scenarios.

Parameter	Black	Hispanic	White	Other	Intervention range
PrEP coverage at baseline (%)a
Coverage	31.9	36.8	51.6	25.8	–
Weekly initiation rateb
Rate	0.0018	0.0020	0.0039	0.0008	1.1×–2× baseline
Proportion of slow stoppers (%)c
Proportion	42.2	42.8	56.0	52.1	+5 to +40 percentage-point
Mean time on PrEP, fast stoppers (weeks)c
Weeks	12.0	13.0	13.9	13.4	–
Mean time on PrEP, slow stoppers (weeks)c
Weeks	66.5	72.2	81.9	73.7	–
Weekly reinitiation rate, fast stoppersc,d
Rate	0.0691	0.0750	0.0664	0.0745	1.25×–5× baseline
Weekly reinitiation rate, slow stoppersc,d
Rate	0.0238	0.0260	0.0283	0.0277	1.25×–5× baseline

Baseline values for PrEP coverage, weekly initiation rate, proportion of slow stoppers, time on PrEP, and weekly reinitiation rates among fast and slow stoppers, shown by race/ethnicity. Intervention ranges describe the multiplicative or additive modifications evaluated in counterfactual scenarios. A full version with footnotes and derivation details is provided as Supplementary Table S1.

^aProportion of PrEP-indicated MSM using PrEP; derived from surveillance.

^bCalibrated to overall PrEP coverage.

^cEstimated from Boston health system EHR data (2012–2023).

^dWeekly rate = multiplicative inverse of the mean weeks between discontinuation and reinitiation, the maximum-likelihood rate under a constant-hazard (exponential) waiting-time assumption within each stopper class. Race/ethnicity was self-identified in the EHR; Black and White categories are non-Hispanic, Hispanic includes any race, and Other combines Asian, American Indian or Alaska Native, Native Hawaiian or Pacific Islander, multiracial, declined/other, missing, or otherwise unclassified records. Rate estimates are point values entered as fixed model inputs; outcome uncertainty is expressed through the 95% simulation intervals rather than through input confidence intervals.

EHR-based parameterization

Race-specific initiation rates, fast/slow-stopper proportions, time to discontinuation, and reinitiation rates were estimated from EHR data from the Electronic medical record Support for Public Health (ESP) system at three Boston-area health systems (2012–2023). The cohort included 37,722 MSM with at least one clinical encounter meeting PrEP indication criteria (Supplementary Figure S1).²⁸ MSM ascertainment required male gender identity, male sex assigned at birth, and documented gay/bisexual identity, male partners, or recent rectal/pharyngeal chlamydia or gonorrhea testing; PrEP indication was assigned at each encounter from sexual behavior, STI history, and clinical notes. The three health systems capture a large and racially and ethnically diverse share of Boston MSM PrEP care but under-represent MSM not engaged in clinical care and those using direct-to-consumer or telehealth PrEP, so the initiation and reinitiation rates are conditional on care engagement. Race/ethnicity was self-identified in the EHR and modeled as Black, Hispanic, White, and Other (Table 1), a marker of structural racism and differential health-system access, not biological risk.

Time on PrEP was defined by continuous prescription coverage. Because ESP captures prescriptions and fills rather than dosing regimen, the stopper classes describe coverage duration rather than daily versus on-demand or long-acting use. Discontinuation was defined as a lapse in coverage beyond a one-week grace period appended to each fill, consistent with prior Boston EHR analyses.⁹ Reinitiation was defined as a new PrEP prescription following a prior discontinuation. The model does not distinguish planned from unplanned discontinuation; seasonal or intermittent users are represented as repeated discontinue–reinitiate cycles governed by the estimated rates rather than as permanent loss, so they are not misclassified as permanently discontinued. Discontinuation and reinitiation rates were estimated as the inverse of the mean time to each event, separately within race and stopper class, while initiation rates were set by calibration to race-specific PrEP coverage; all were passed directly into the simulation as race- and class-specific transition probabilities. Baseline race-specific values are summarized in Table 1 and Supplementary Table S1, and analytic-cohort characteristics in Supplementary Table S6. Full derivation, cohort flow, and a comparison of fast/slow cut-points appear in Appendix Section 7.6.

Ethics

The ARTnet study and the analysis of Boston health-system electronic health record data were approved by the Emory University Institutional Review Board (protocol [IRB 00003240]). ARTnet participants provided electronic informed consent. The electronic health record analysis used retrospective, routinely collected clinical data, for which the Institutional Review Board granted a waiver of informed consent.

Intervention scenarios

We simulated 10-year interventions on the three continuum levers from the calibrated 2024 model state. Uptake interventions scaled race-specific baseline initiation rates by 1.1×–2.0×. Persistence interventions shifted a fraction of fast stoppers into the slow-stopper category by +5 to +40 percentage points. Reinitiation interventions scaled race- and stopper-class-specific baseline rates by 1.25×–5.0×, mainly displaying 1.5x–5.0x. Joint heatmaps used narrower visualization grids (uptake 1×–2×, persistence +0.1 to +0.4, reinitiation 1×–3×). These ranges characterize the response surface of each lever rather than mapping any specific intervention; lower bounds represent modest single-component improvements and upper bounds are deliberately idealistic scenarios, counterfactual intensities included to map the limit of each lever's potential rather than to represent a currently achievable program.

Each intervention was simulated under two implementation strategies. In the equal-implementation strategy, the intervention was applied uniformly to all MSM (a population-wide program). The same multiplicative change (uptake, reinitiation) or additive percentage-point change (persistence) was applied to every race/ethnicity group's baseline rate. In the equity-focused strategy, the intervention was applied only to Black and Hispanic MSM, representing a tailored program. Matched scenarios set Black/Hispanic slow-stopper proportions or reinitiation rates to corresponding White baselines; because some Black/Hispanic reinitiation rates exceeded White rates, matching was not always a scale-up. For interpretability, the main text emphasizes comparison intensities (1.5×, 2×, and 5×) at which the three levers can be contrasted head-to-head. Comparison intensities are multiples of the baseline reinitiation rate; at each, uptake and persistence are evaluated at their sweep maxima for the head-to-head comparison. The full scenario sweep is reported in Supplementary Tables S2–S5.

Statistical analysis

Primary outcomes were year-10 PrEP coverage, cumulative HIV incidence per 100 person-years (PY), percent of HIV infections averted (PIA) versus baseline, and number needed to treat (NNT) to avert one infection. NNT was calculated as the additional cumulative person-years on PrEP delivered under the intervention, relative to baseline, divided by the number of HIV infections averted versus baseline, so that numerator and denominator are both differences against the same reference; lower values indicate greater prevention efficiency. Race/ethnicity-specific outcomes were incidence, relative incidence ratios (Black-to-White, Hispanic-to-White), and absolute disparities (Black-White, Hispanic-White). All outcomes are reported as medians with 95% simulation intervals, taken as the 2.5th and 97.5th percentiles of the replicate distribution and capturing stochastic variability across model replicates rather than parameter uncertainty. Each scenario was simulated in at least 64 replicates from the calibrated model state, and up to 120 per cell in the joint uptake-by-reinitiation grid. No hypothesis testing was performed and no sample-size calculation applies to a simulation study of this design.

Role of the funding source

The funders had no role in study design, data analysis or interpretation, manuscript preparation, or the decision to submit for publication. The corresponding author had full access to all data and final responsibility for the decision to submit for publication.

Results

At baseline, overall PrEP coverage among PrEP-indicated MSM was 41.5% (16,063 of 38,734) and overall HIV incidence was 0.95 per 100 person-years (PY) (95% SI 0.87–1.03). Over the simulated 2024–2034 horizon the baseline model produced 9352 cumulative infections (95% SI 9038–9611), the denominator for all percentages of infections averted below. Baseline HIV incidence was substantially higher among Black (1.47 per 100 PY, 95% SI 1.27–1.70) and Hispanic (1.94 per 100 PY, 95% SI 1.78–2.23) MSM than among White MSM (0.60 per 100 PY, 95% SI 0.52–0.68).

We contrasted the three continuum levers, uptake, persistence, and reinitiation, at the comparison intensities (Table 2, Fig. 2). At a 2x scale-up, reinitiation averted 15.8% of infections over 10 years (1481 infections; 95% SI 13.3–19.8; NNT 40), compared with 10.0% for 2× uptake (939 infections; 95% SI 6.9–12.7; NNT 29) and 5.6% for the maximum persistence scenario (+40 percentage-point slow stoppers; 524 infections; 95% SI 2.8–8.1; NNT 53). A 1.5× reinitiation scale-up averted 10.0% of infections (940 infections; 95% SI 7.5–12.4; NNT 37), matching the PIA of a 2× increase in initiation. At its maximum modeled intensity (5×), reinitiation averted 32.6% of infections (3045 infections; 95% SI 29.6–35.6; NNT 40), more than triple the maximum uptake scenario and nearly six times the maximum persistence scenario. Reinitiation's NNT was stable near 40 across all comparison intensities, between uptake (29) and persistence (53), so its advantage lay in the scale of benefit attainable rather than in efficiency per person-year of PrEP delivered. Persistence was the weakest lever, and at lower intensities its effect was not distinguishable from no effect: a +20 percentage-point increase in slow stoppers averted 2.2% of infections (208 infections; 95% SI −0.4 to 4.8).

Table 2.

Projected impact of PrEP uptake, persistence, and reinitiation scenarios on PrEP coverage, HIV incidence, percent of infections averted, and number needed to treat, among MSM in Boston over 10 years.

Scenario	PrEP coverage (%)	HIV incidence (per 100 PY)	PIAa (%)	Infections averted, n	NNTb
Baseline	41.5 (41.1, 41.8)	0.95 (0.87, 1.03)	–	–	–
Equal-implementation scenarios
1.5× initiation	45.9 (45.5, 46.2)	0.87 (0.79, 0.93)	5.9 (3.6, 8.8)	550 (333, 826)	29 (20, 45)
2× initiation	48.8 (48.4, 49.1)	0.82 (0.74, 0.90)	10.0 (6.9, 12.7)	939 (644, 1190)	29 (23, 42)
+20 pp slow stoppers	43.9 (43.6, 44.4)	0.93 (0.83, 1.01)	2.2 (−0.4, 4.8)	208 (−39, 445)	56 (−325, 1070)
+40 pp slow stoppers (max)	46.5 (46.1, 46.9)	0.91 (0.83, 0.98)	5.6 (2.8, 8.1)	524 (260, 754)	53 (38, 101)
1.5× reinitiation	49.5 (48.9, 50.0)	0.85 (0.77, 0.92)	10.0 (7.5, 12.4)	940 (705, 1163)	37 (30, 50)
2× reinitiation	54.8 (54.3, 55.1)	0.78 (0.72, 0.85)	15.8 (13.3, 19.8)	1481 (1247, 1856)	40 (32, 46)
5× reinitiation (max)	68.3 (67.9, 68.7)	0.61 (0.57, 0.67)	32.6 (29.6, 35.6)	3045 (2765, 3328)	40 (36, 43)
Equity-focused scenarios (Black and Hispanic MSM only)
2× initiation (B/H)	44.2 (43.8, 44.6)	0.87 (0.80, 0.97)	6.5 (3.8, 9.6)	607 (357, 897)	18 (12, 28)
Matched persistence	42.8 (42.4, 43.2)	0.93 (0.87, 1.00)	2.1 (−1.0, 4.8)	193 (−93, 447)	28 (−174, 165)
2× reinit fast (B/H)	43.8 (43.4, 44.1)	0.89 (0.84, 0.97)	5.0 (2.3, 8.0)	468 (214, 749)	25 (16, 53)
Matched reinitiation	42.7 (42.3, 43.1)	0.93 (0.85, 1.00)	1.6 (−1.1, 5.7)	153 (−107, 530)	22 (−95, 704)

Rows show the baseline and selected comparison intensities for equal-implementation and equity-focused implementations. Full scenario sweeps (1.1×–2.0× uptake, +5 to +40 percentage-point persistence, 1.5×–5× reinitiation for overall, fast-stopper-only, and slow-stopper-only) are in Supplementary Table S2 (equal-implementation) and Supplementary Table S4 (equity-focused).

^aPIA: percent of HIV infections averted over 10 years versus baseline.

^bNNT: additional person-years on PrEP delivered by the intervention, relative to baseline, per HIV infection averted. Values are medians (95% simulation intervals). NNT intervals are retained even when the PIA interval crosses zero; negative or very wide bounds indicate unstable interval estimates. Matched scenarios set Black/Hispanic parameter values to corresponding White baselines and were not necessarily scale-ups.

Fig. 2.

Joint-intervention heatmap showing (A) percent of HIV infections averted and (B) number needed to treat (additional person-years on PrEP per infection averted) for scenarios varying PrEP uptake (x-axis; 1×–2× baseline initiation) and PrEP reinitiation (y-axis; 1×–3× baseline reinitiation) simultaneously. The bottom row (1× reinitiation) corresponds to uptake-only interventions; the left column (1× initiation) corresponds to reinitiation-only interventions; the interior cells show the combined effect of simultaneously scaling both levers. The main-effects comparison of the three continuum levers (reinitiation > uptake > persistence) is visible along the edges of the figure; interior cells show that uptake-and-reinitiation interventions produce approximately additive combined gains. This heatmap shows the continuous uptake-by-reinitiation response surface over the visualization grid; the persistence-by-reinitiation surface is shown in Supplementary Figure S3, and the discrete comparison intensities, including 5× reinitiation, are reported in Table 2.

Fig. 2 displays the joint scale-up of uptake and reinitiation as a heatmap of 10-year PIA. The interior cells show that combined uptake-and-reinitiation interventions produced approximately additive gains. Joint PrEP coverage is shown in Supplementary Figure S2, the persistence-by-reinitiation PIA and NNT heatmap in Supplementary Figure S3, and the single-lever main effects in Supplementary Figure S4.

Subgroup scale-ups identified short initial PrEP episodes (under 27 weeks) as a modestly higher-impact restart context than longer episodes. Restricted to short-episode discontinuation, PIA ranged from 5.0% (464 infections) at 1.5× to 19.7% (1841 infections) at 5×; under longer-episode discontinuation, from 5.0% (464 infections) to 15.5% (1449 infections; Supplementary Table S2). Both subgroup-only scenarios remained below the corresponding all-stopper benefit at every modeled intensity. Supplementary Figure S5 shows the PrEP-engagement composition of the PrEP-eligible population over the horizon, and Supplementary Table S7 the year-10 composition of its off-PrEP portion by prior PrEP use.

Population-wide interventions reduced overall and race-specific HIV incidence (Table 3). At 2× uptake, overall incidence fell from 0.95 to 0.82 per 100 PY, while the Black-to-White and Hispanic-to-White ratios were essentially unchanged (2.51–2.55 and 3.26–3.25). Absolute disparities generally narrowed under uptake and reinitiation, but relative disparities hovered near baseline or widened, particularly at high-intensity persistence and reinitiation.

Table 3.

Race/ethnicity-specific HIV incidence and relative and absolute disparity outcomes for the comparison scenarios in Table 2, with White MSM as the reference group.

Scenario	Black incidence	Hispanic incidence	White incidence	B/W ratioa	H/W ratioa	B-W absoluteb	H-W absoluteb
Baseline	1.47 (1.27, 1.70)	1.94 (1.78, 2.23)	0.60 (0.52, 0.68)	2.51 (2.06, 2.96)	3.26 (2.86, 3.89)	0.88 (0.65, 1.10)	1.35 (1.17, 1.63)
Equal-implementation scenarios
1.5× initiation	1.34 (1.16, 1.55)	1.76 (1.51, 2.12)	0.55 (0.48, 0.62)	2.44 (2.03, 2.94)	3.20 (2.65, 3.96)	0.80 (0.60, 1.01)	1.22 (0.93, 1.57)
2× initiation	1.31 (1.14, 1.48)	1.66 (1.38, 1.97)	0.51 (0.45, 0.60)	2.55 (2.12, 3.12)	3.25 (2.44, 4.07)	0.79 (0.62, 0.96)	1.15 (0.81, 1.48)
+20 pp slow stoppers	1.47 (1.23, 1.65)	1.94 (1.70, 2.20)	0.58 (0.51, 0.66)	2.57 (2.11, 2.87)	3.34 (2.97, 4.06)	0.89 (0.67, 1.04)	1.35 (1.14, 1.64)
+40 pp slow stoppers (max)	1.45 (1.26, 1.64)	1.89 (1.57, 2.12)	0.54 (0.50, 0.62)	2.61 (2.16, 3.20)	3.45 (2.65, 3.99)	0.90 (0.69, 1.11)	1.34 (1.00, 1.57)
1.5× reinitiation	1.36 (1.16, 1.57)	1.73 (1.52, 2.03)	0.51 (0.45, 0.58)	2.67 (2.14, 3.21)	3.51 (2.85, 4.28)	0.86 (0.63, 1.06)	1.23 (1.00, 1.54)
2× reinitiation	1.24 (1.06, 1.47)	1.66 (1.43, 1.84)	0.46 (0.40, 0.55)	2.63 (2.06, 3.25)	3.53 (2.90, 4.36)	0.77 (0.55, 1.01)	1.21 (0.94, 1.37)
5× reinitiation (max)	0.97 (0.85, 1.10)	1.30 (1.08, 1.53)	0.34 (0.30, 0.40)	2.80 (2.32, 3.39)	3.78 (2.87, 4.68)	0.62 (0.49, 0.75)	0.95 (0.71, 1.18)
Equity-focused scenarios (Black and Hispanic MSM only)
2× initiation (B/H)	1.32 (1.13, 1.52)	1.82 (1.57, 2.11)	0.58 (0.50, 0.66)	2.46 (1.89, 2.90)	3.15 (2.65, 3.88)	0.77 (0.56, 1.00)	1.25 (0.97, 1.56)
Matched persistence	1.44 (1.22, 1.65)	1.90 (1.64, 2.17)	0.60 (0.52, 0.68)	2.43 (2.05, 2.86)	3.23 (2.68, 3.84)	0.86 (0.65, 1.08)	1.32 (1.08, 1.60)
2× reinit fast (B/H)	1.38 (1.17, 1.58)	1.79 (1.55, 2.04)	0.58 (0.50, 0.67)	2.40 (1.95, 2.87)	3.11 (2.55, 3.77)	0.82 (0.60, 1.04)	1.23 (0.99, 1.50)
Matched reinitiation	1.45 (1.24, 1.68)	1.91 (1.66, 2.20)	0.60 (0.52, 0.68)	2.41 (1.92, 2.84)	3.21 (2.66, 3.86)	0.85 (0.63, 1.07)	1.32 (1.06, 1.59)

Population-wide interventions reduced overall and race-specific incidence; absolute disparities generally narrowed, while relative disparities hovered near baseline or widened. Equity-focused interventions reduced incidence among Black and Hispanic MSM with more favorable relative-disparity patterns, though substantial inequities remained. The full equal-implementation sweep is in Supplementary Table S3 and the full equity-focused sweep is in Supplementary Table S5.

^aRelative disparity = race-specific incidence/White incidence.

^bAbsolute disparity = race-specific incidence – White incidence (per 100 PY). HIV incidence is per 100 person-years; values are medians (95% simulation intervals). Matched scenarios set Black/Hispanic parameter values to corresponding White baselines and were not necessarily scale-ups.

Equity-focused interventions had smaller population-level PIA than population-wide counterparts but more favorable relative-disparity patterns. A 2× uptake scale-up tailored to Black and Hispanic MSM averted 6.5% overall (607 infections) and modestly narrowed both incidence ratios relative to baseline (Table 3). In the full equity-focused sweep (Supplementary Table S5), reinitiation scale-ups produced the most favorable relative disparity ranges among the three equity-focused levers, with Black-to-White ratios of 2.19–2.53 and Hispanic-to-White ratios of 2.92–3.28 across reinitiation intensities (versus baseline ratios of 2.51 and 3.26), although there was no clear monotonic dose–response with increasing reinitiation intensity, with the ratio ranges overlapping within their simulation intervals. Across all equity-focused scenarios, Black and Hispanic MSM continued to experience HIV incidence well above White MSM, indicating that PrEP continuum interventions alone were insufficient to close the gap.

Discussion

In a race/ethnicity-stratified network model of Boston MSM parameterized with health-system EHR data, interventions scaling PrEP reinitiation averted substantially more HIV infections than interventions scaling initial uptake or persistence. When scaled equally, no lever reliably reduced relative racial and ethnic disparities, although absolute disparities often narrowed. To our knowledge, this is the first modeling study to represent reinitiation as a distinct implementation target and to compare it directly against uptake and persistence. The field's focus on initial uptake, appropriate during early scale-up, increasingly understates the intervention space in mature PrEP environments.

The mechanism behind the reinitiation advantage depends on a mature PrEP environment. In a decade-old PrEP environment, nearly half of PrEP-eligible individuals who are off PrEP are prior users rather than PrEP-naive (48.0% at baseline; Supplementary Table S7). Prior users also face lower structural and informational barriers to restart than people new to PrEP.^14,15 In our EHR data, weekly reinitiation rates among former users were an order of magnitude higher than baseline initiation rates among PrEP-naive individuals. A multiplicative increase on reinitiation therefore yields a larger absolute change in starts than the same multiplier on initiation. This advantage attenuates where cumulative PrEP uptake, and thus the accumulated pool of former users, is small. The finding that short initial episodes were modestly higher-impact than longer ones is counterintuitive but interpretable. Short episodes were more common, had higher baseline reinitiation rates, and responded more steeply to reduced re-engagement friction; subgroup-only scale-ups still produced smaller benefits than all-stopper scale-ups at every intensity. Notably, Black and Hispanic MSM with short episodes reinitiated at higher weekly rates than White MSM (Table 1), inverting the initial-uptake pattern. The modeled response also plausibly underlies the elevated post-discontinuation HIV incidence reported by Spinelli and colleagues.¹¹ Reinitiation interventions target this elevated post-stop person-time directly, and so yield the largest projected return.

The contrast between population-wide and equity-focused implementations exposes a trap in continuum-based PrEP planning. Population-wide improvements on any of the three continuum levers reduced absolute HIV incidence but did not reduce (and in some scenarios widened) relative racial and ethnic disparities. This is not an artifact: intervention benefits accrue in proportion to the eligible pool, which is larger among White MSM at baseline. The modeled disparities also arise from differences in baseline HIV prevalence and from assortative, race/ethnicity-homophilous sexual network structure, not from PrEP-continuum behavior alone. The same structural dynamic was described for initiation alone by Jenness and colleagues in a 2018 modeling study of PrEP and racial and ethnic disparities.¹⁸ Our results extend this to persistence and reinitiation, general across continuum levers.

By contrast, equity-focused implementations reduced incidence among Black and Hispanic MSM and produced more favorable disparity patterns, although the dose–response was not monotonic. This non-monotonicity reflects the sensitivity of the relative ratios to small movements in the low-incidence White denominator and the indirect lowering of White incidence through the shared transmission network, rather than a substantive reversal, and the ranges overlap within their simulation intervals. Equity-focused reinitiation produced the most favorable relative-disparity ranges of the three levers in the full equity-focused sweep (Supplementary Table S5). Because equity-focused scenarios deliver less total PrEP than population-wide scenarios yet still reduce Black and Hispanic incidence, their higher per-unit targeting efficiency (NNT 25 versus 40 for the corresponding population-wide scenario), rather than smaller scale alone, drives the more favorable patterns. These disparity conclusions are sensitive to whether disparities are measured in relative or absolute terms, which we report jointly (Table 3). Even under the most aggressive equity-focused scenarios, substantial inequities persisted. PrEP continuum interventions alone cannot close the gap. Structural drivers the model omits (health-system access, insurance continuity, medical mistrust, and structural racism in care) must be addressed in parallel.

These findings have implications for PrEP intervention design. First, reinitiation interventions are operationally distinct from generic uptake campaigns: refill-gap registries, EHR alerts, portal or pharmacy outreach, rapid-restart protocols, and telehealth visits can all be triggered by prior PrEP use. EHR-triggered prompts already increase PrEP prescribing in PrEP-naive users,²⁹ and the same infrastructure can be repurposed for restart. Second, structural friction in PrEP re-enrollment (insurance authorization, baseline lab testing, pharmacy pickup) should be evaluated and reduced where clinically appropriate. Pharmacist-led PrEP services under collaborative practice agreements are one scalable model for reducing these friction points.³⁰ Even modest reductions in re-engagement friction could yield population-level gains exceeding comparable uptake interventions. Third, the restart encounter is an under-used decision point for modality choice: a user returning after a daily-oral interruption may prefer injectable or event-driven PrEP. Fourth, because the most favorable relative-disparity ranges in our analysis came from equity-focused reinitiation specifically, intervention programs that combine reinitiation focus with geographic and demographic tailoring to Black and Hispanic MSM offer a distinct route to both incidence and equity gains, unavailable through population-wide implementation of any single lever.

Our study has several limitations. First, the model is calibrated to Boston-specific HIV and PrEP parameters, and the EHR cohort reflects care-engaged MSM. Generalizability to other settings, to MSM not in clinical care, and to direct-to-consumer PrEP users outside these systems is uncertain; initiation and reinitiation rates are conditional on care engagement and EHR documentation of these characteristics. Race/ethnicity was used as a marker of structural racism and health-system access, not biological risk; residual confounding and within-group heterogeneity remain. The modeled population is cisgender MSM by ascertainment, and the findings do not extend to transgender people. The reinitiation advantage is therefore most directly generalizable to populations already linked to PrEP-capable care, although the ordering (reinitiation > uptake > persistence) is likely robust to baseline shifts. As partial external validation, the model reproduces independent surveillance not used as calibration targets, including HIV diagnostic delay consistent with national projections and baseline race/ethnicity-specific incidence consistent with Massachusetts surveillance, although full out-of-sample validation is inherently limited for any context-specific calibrated model. Second, we did not represent long-acting injectable PrEP or event-driven dosing, both likely to change reinitiation dynamics. Injectable PrEP may alter short-episode discontinuation; this is a priority for future modeling. Third, the two-class dichotomy simplifies a continuous duration distribution but improves on the memoryless restart of prior models. Fourth, our equity-focused scenarios assume tailored interventions reach Black and Hispanic MSM at the specified rates without attrition; real-world effectiveness may be lower. Fifth, the EHR-derived rates enter the model as fixed point estimates rather than distributions, so a probabilistic sensitivity analysis propagating their uncertainty is a priority for future work, although the qualitative lever ordering is stable across the intensity sweep already reported. Sixth, the Other category aggregates heterogeneous subpopulations, so its parameters represent an average across distinct groups rather than a coherent category, and it is not part of the Black, Hispanic, and White disparity comparison. Seventh, stopper class is a fixed agent attribute, whereas an individual's discontinuation pattern may in reality evolve across episodes. Finally, our equity-focused scenarios do not hold total PrEP volume constant; a fixed-total-PrEP reallocation experiment is an informative complementary design for future work.

In conclusion, PrEP reinitiation is an under-recognized, disproportionately impactful analog to care re-engagement. Reinitiation interventions averted more infections than uptake or persistence interventions, at a stable NNT near 40 additional person-years on PrEP per infection averted, and this advantage was consistent across scale-up intensities. Equity-focused reinitiation interventions offer a distinct route to both incidence reduction and disparity reduction, though they cannot close the gap alone and must be paired with structural interventions addressing the drivers of disparity. Prospective evaluation of reinitiation-specific strategies, such as EHR-triggered prompts at return visits and low-friction re-enrollment, is warranted.

Contributors

CC led the design, analysis, and writing of the paper. SMJ supervised the project and contributed to the design, analysis, writing, and editing of the paper. ALG contributed to the design and analysis of the paper. HKG, KHM, and PS led the implementation of the parent EHR data at the Boston-area health systems and contributed to the design and editing of the paper. JLM, JJ, and NDC contributed to the design, writing, and editing of the paper. CC and SMJ accessed and verified the underlying data. SMJ (the corresponding author) was responsible for the decision to submit the manuscript. All authors read and approved the final version of the manuscript.

Data sharing statement

The simulation code (EpiModelHIV) is available at https://github.com/EpiModel/EpiModelHIV. Calibration scripts, intervention scenario definitions, and analysis code specific to this study are available at the project repository (URL to be added at acceptance) and will be deposited in Zenodo with a permanent DOI at publication. The Boston EHR cohort data underlying parameter estimation are governed by data use agreements with the participating health systems and are not publicly available; aggregate-level race- and stopper-class-specific estimates used in the model are reported in Table 1 and Supplementary Table S1, and are sufficient to reproduce all simulated intervention scenarios and outcomes.

Declaration of interests

SMJ has received a grant from Merck paid to his institution and consulting fees from Gilead Sciences and the Elton John AIDS Foundation. JLM received a one-time honorarium from MyBodyPro for developing medical education content. JJ receives salary support from a grant paid to his institution by ViiV Healthcare. HKG has received consulting fees from Reach Capital and holds unpaid leadership roles as board chair and co-founder of Shift Collaborative and as an advisory board member for the Arizona State University Substance Use Health Record Sharing (SHARES) initiative. KHM has received grants paid to his institution from Merck, Gilead Sciences, and ViiV Healthcare, and has served on the scientific advisory boards of Merck, Gilead Sciences, and ViiV Healthcare. PS has received grants paid to his institution from the National Institutes of Health (K01AI167733) and a Department of Medicine career investment award from Boston University Chobanian and Avedisian School of Medicine. All other authors declare no competing interests.