A $19-a-Month App Cut Detectable HIV Viral Load 58 Percent. By Month Twelve the Effect Was Gone.

The trial was designed to solve a problem that software is unusually well suited to: men who are living with HIV, who use methamphetamine or another stimulant heavily enough to meet a diagnostic threshold, and who keep falling out of care. It enrolled 286 of them across the continental United States without a single clinic visit. Blood was drawn remotely. The intervention was an app.

At six months, it worked. Men randomized to the START application had 58 percent lower odds of a detectable HIV viral load than men given a website of care referrals — an adjusted odds ratio of 0.42, with a confidence interval of 0.19 to 0.93. By twelve months, there was no difference between the groups on viral load, on virologic rebound, or on adherence. The effect did not shrink. It disappeared.

The results were published in eClinicalMedicine, the Lancet Discovery Science journal, in a paper led by Adam W. Carrico and a long author list that includes Michaela Larson, Leah Davis-Ewart, Delaram Ghanooni and Karin Horvath.

What the trial actually tested

START — Supporting Treatment Adherence for Resilience and Thriving — is not a reminder system. It is built on the information-motivation-behavioural skills model, with a second layer borrowed from Revised Stress and Coping Theory: the idea that gratitude, mindfulness and other positive psychological processes do work during chronic stress, and that adherence is a behaviour people can be taught rather than nagged into.

Concretely, the intervention arm got six modules containing 31 on-demand videos covering ten positive psychological intervention skills, plus eight audio meditation files. There were brief check-ins for self-monitoring adherence, substance use and mood, with graphs so a participant could see their own trend lines. The control arm got a website with links to crisis counselling, mental health care, substance use treatment and HIV care. Both are reasonable things to hand someone. Only one includes structured skills training.

Eligibility was narrow in a way that matters for how the results should be read. Participants had to be sexual minority men, 18 or older, diagnosed with HIV at least three months earlier, currently on antiretroviral therapy — and they had to clear at least one risk marker: a detectable viral load in the past year, disengagement from care, or adherence below 90 percent. On top of that, they had to screen positive for moderate or severe stimulant use disorder on the ASSIST.

That combination is deliberate. Up to one in three sexual minority men who acquire HIV in the United States report persistent methamphetamine use, and the clinical consequence is not ambiguous: worse adherence, more detectable viral load, more onward transmission risk. The trial was aimed at the exact seam where the HIV epidemic and the stimulant epidemic are stitched together.

The number nobody quoted

The relative figure — 58 percent — is the one that travels. The absolute figure is more useful to anyone deciding whether to fund this.

The risk difference was minus 0.21, which puts the number needed to treat at roughly 4.75. In plain terms: for every five men given the app instead of the referral website, one fewer had a detectable viral load six months later. For a digital intervention with no clinical staff attached, that is a large effect, and it is the strongest argument in the paper.

It also has an expiry date. The NNT of 4.75 is a six-month number. Nothing in the twelve-month data supports a comparable claim, because at twelve months the arms converged.

Rebound, not rescue

The second thing the headline hides is where the benefit came from. It came almost entirely from preventing virologic rebound, not from rescuing people in free fall.

Among the subset of participants who entered the trial with an undetectable viral load, the odds of rebound at six months were 75 percent lower in the START arm. In the control arm, nearly half of those men rebounded. In the intervention arm, about a quarter did. The rebounds that were prevented clustered at low-level viraemia — 301 to 838 copies per millilitre — which is the range that precedes the louder failures.

That is a maintenance effect, and it reframes who the app is for. START is not a re-engagement tool for the man who stopped taking his medication eight months ago and stopped answering his phone. It is a tool for the man who is suppressed right now and is one bad stretch away from not being suppressed. Those are different clinical problems with different price tags, and only one of them is being solved here.

The screener was also a filter

There is a selection story inside the enrollment numbers. Of 2,346 people who screened as eligible, 508 enrolled. Of those, 286 were randomized. Roughly half of the enrolled pool never made it to randomization, and the paper attributes most of that attrition to one thing: they could not produce a viable remote blood sample for the baseline viral load.

That is a technical failure with a social shape. Consenting to a trial, receiving a kit, and completing a remote phlebotomy protocol requires a stable address, a working phone, a few uninterrupted minutes and enough executive function to follow written instructions. Those are precisely the capacities that stimulant use disorder erodes and that sustained engagement in HIV care depends on. So the trial's population is, by construction, the more organized slice of the population it wants to help — and the app's measured effect is the effect in people who could already complete a blood draw.

The authors do not oversell this. They name it as a limitation and call for better coordination of remote phlebotomy in future studies.

The cliff at six months

The most uncomfortable finding in the paper is not statistical. It is a line about check-ins.

Participants in the intervention arm completed a median of 11 check-ins during the first six months. After the six-month mark, the median dropped to zero. Engagement did not taper. It stopped.

And the effect stopped with it. The window in which the app demonstrably worked is the same window in which people were using it.

That correlation invites an obvious reading — the app worked because it was used, and it stopped working because it stopped being used — but the trial cannot cleanly support it, because something else also stopped at six months. During the first half of the study, participants who had not logged in received weekly text message reminders from study staff. That staffing cost is included in the program budget. After six months, it was not.

So there are two candidate explanations for the cliff, and the design cannot separate them: the software lost its effect, or the human scaffolding around the software was withdrawn. A field that keeps publishing digital interventions with six-month effects would benefit enormously from knowing which.

Why 300 copies per millilitre matters

The assay used to define the primary outcome had a lower limit of detection of 300 copies per millilitre. The standard clinical definition of viral suppression is below 200.

This cuts in two directions, and both should be stated. A coarser threshold means the trial could not see small differences between arms, so the measured effect is if anything conservative — the app may have prevented blips that the assay was blind to. But it also means the outcome is not the outcome clinicians use when they tell a patient they are undetectable, and the 58 percent figure should not be translated directly into a clinical promise.

The authors note the constraint themselves and flag objective adherence measures — urine or hair toxicology — as the next step. Self-reported adherence at six months showed a nearly 2.5-fold advantage for the app, but at a p-value of 0.051. Marginal is the right word. It is the kind of result that becomes a talking point in a press release and a caveat in a journal club.

What $19 a month buys

The cost analysis is unusually concrete for a digital health trial. Delivering START for twelve months cost $41,002 including fixed start-up and reprogramming, which works out to $231 per participant per year, or about $19 per month across the 143 men randomized to the intervention arm.

For context, that is less than a single specialist visit in most markets, and roughly the price of a monthly transit pass in a mid-sized city. The trial also declined to make the intervention easier than it would be in the real world: participants were not given phones, data plans or financial incentives for completing modules. The median system usability score was 70, which is acceptable rather than delightful.

Cheap, however, is doing a lot of work in that paragraph. The $19 figure covers software and reminders. It does not cover the thing the authors recommend adding next.

The payment problem the paper points at

In their discussion, Carrico and colleagues suggest two ways to make the benefit last: pair the app with remote contingency management, the incentive-based approach with the strongest trial evidence for stimulant use disorder, and add individually delivered telehealth counselling targeting adherence or substance use.

Both are good suggestions. Both run into the same wall. Contingency management works and is chronically under-reimbursed — there are few billing codes, and the incentives themselves are the kind of line item that procurement officers query. Telehealth counselling requires clinicians, which is the workforce problem digital tools were supposed to route around.

Which leaves the honest summary of this trial: a $19-a-month app prevented a meaningful share of early virologic rebound in a population at real risk of it, for exactly as long as people kept opening it, and the effect vanished when the check-ins and the reminders did. That is a genuine result, not a null one. It is also a result about maintenance, delivered to a field that keeps hoping software can do the harder job of re-engagement.

For men living with HIV and a stimulant use disorder, the practical takeaway is not that an app replaces care. It is that the dose of support matters as much as the medium, and that the intervention with the best evidence for the stimulant side of the problem — contingency management — remains the one hardest to get paid for.

The trial was conducted entirely remotely between June 2022 and May 2024. The paper appears in eClinicalMedicine volume 101.