Design: Five-year longitudinal survival analysis with Kaplan-Meier estimates, Cox regression and random forest classification
Population: 865 mental health apps listed in a curated index, followed over 2,142 days
Primary outcome: Removal from the database as a marker of app attrition
Effect: 53.6 per cent removed; Android-only apps hazard ratio 2.32; Cox C-index 0.77, random forest AUC 0.82; sleep apps most durable, smoking cessation least

Eight hundred and sixty-five mental health apps listed in a curated database were followed for 2,142 days with Kaplan-Meier survival estimates and Cox regression across 54 clinical, technical and commercial predictors. Four hundred and sixty-four, 53.6 per cent, were removed during the study period.

Attrition was not random. Sleep apps lasted longest; smoking cessation apps had the highest risk of disappearing, aside from a small group of apps for schizophrenia. The single most prominent predictor was platform exclusivity: an app available on only one operating system was far more likely to go, and Android-only apps carried more than double the hazard (HR 2.32). The models discriminated reasonably - Cox C-index 0.77, random forest area under the curve 0.82.

This changes how a recommendation should be made. Advising an app is a treatment decision with a follow-up implication, and an app that disappears is a treatment that stops without anyone deciding to stop it - potentially at a bad moment, since these are patients whose engagement is fragile. Before recommending one, check that it exists on both platforms, that it has been updated recently, and that the patient knows what to do if it goes. In India, where Android dominates and Android-only apps are exactly what a patient is most likely to find, the highest-risk category is also the most available - which is a reason to check rather than to despair, and a reason to keep a paper or clinician-held alternative for anything the treatment plan depends on.

Check an app exists on both iOS and Android before recommending it; single-platform apps carried over twice the hazard.
Look at the date of the last update, not just the reviews.
Tell the patient what to do if the app disappears, and record which app you recommended.
Do not let an app carry an essential part of the treatment plan - safety plans especially.
Note the caveat: this measures removal from a curated database, not necessarily removal from every app store.

The statistics, in plain English

A hazard ratio of 2.32 describes the rate of removal over time rather than a simple proportion, so Android-only apps were not twice as likely to be gone at the end but disappearing at roughly twice the rate throughout. A C-index of 0.77 and an area under the curve of 0.82 mean the models identify at-risk apps usefully but not reliably enough to predict any individual one. The outcome is removal from a curated database, which is a proxy for disappearance and may also reflect the curators' inclusion decisions.