Background: To reach South Africa's targets for HIV treatment and viral suppression, retention on antiretroviral therapy (ART) must increase. Much effort and resources have been invested in tracing those already disengaged and returning them to care programs with mixed success. Here we aim to successfully identify ART clients at risk of loss from care prior to disengagement. Methods and Findings: We applied a previously developed machine learning and predictive modelling algorithm (PREDICT) to routinely collected ART client data from the SLATE I and SLATE II trials, which evaluated same-day ART initiation in 2017-18. Using a primary outcome of an interruption in treatment (IIT), defined as missing the next scheduled clinic visit by >28 days, we investigated the reproducibility of PREDICT in SLATE datasets. We also tested two risk triaging approaches: 1) threshold approach classifying individuals into low, moderate, or high risk of IIT; and 2) archetype approach identifying subgroups with characteristics associated with risk of ITT. We report associations between risk category groups and subsequent IIT at the next scheduled visit using crude risk differences and relative risks with 95% confidence intervals. SLATE datasets included 7,199 client visits for 1,193 clients over 14 months of follow-up. The algorithm achieved 63% accuracy, 89% negative predictive value, and an area under the curve of 0.61 for attendance at next scheduled visit, similar to previous results using only medical record data. The threshold approach consistently and accurately assigned levels of IIT risk for multiple stages of the care cascade. The archetype approach identified several subgroups at increased risk of IIT, including those late to previous appointments, those returning after a period of disengagement, those living alone or without a treatment supporter. Behavioural elements of the archetypes tended to drive risk of treatment interruption more consistently than demographics; e.g. adolescent boys/young men who attended visits on time experienced lowest rates of treatment interruption (10%, PREDICT datasets and 7% SLATE datasets), while adolescent boys/young men returning after previously disengaging from care had highest rates of subsequent treatment interruption (31%, PREDICT datasets and 40% SLATE datasets). Conclusion: Routinely collected medical record data can be combined with basic demographic and socioeconomic data to assess individual risk of future treatment disengagement using machine learning and predictive modelling. This approach offers an opportunity to intervene prior to and potentially prevent disengagement from HIV care, rather than responding only after it has occurred.

The authors have declared no competing interest.

This study has been made possible by the generous support of the American People and the President's Emergency Plan for AIDS Relief (PEPFAR) through the United States Agency for International Development (USAID), including bilateral support through USAID South Africa's Accelerating Program Achievements to Control the Epidemic under the terms of cooperative agreement 72067418CA00029 to HE2RO and cooperative Agreement 72067419CA00004. SR and MM are additionally supported by the Bill and Melinda Gates Foundation (Grant Number: INV-031690). The contents are the responsibility of the authors and do not necessarily reflect the views of PEPFAR, USAID or the United States Government. The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The corresponding author had final responsibility for the decision to submit for publication.

I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.

Yes

The details of the IRB/oversight body that provided approval or exemption for the research described are given below:

All analyses of de-identified data from human subjects were approved by and carried out in accordance with relevant guidelines and regulations as set out by the Human Research Ethics Committee of the University of the Witwatersrand (Medical). This study involved secondary analysis of two data sources: 1) deidentified data collected as part of routine care, for which the requirement for individual patient consent was waived by the Human Research Ethics Committee of the University of the Witwatersrand for protocols M140201 and M210472 during the study approval; and 2) de-identified clinical trial collected as part of the SLATE I and SLATE II trials (Clinicaltrials.gov registration NCT02891135). Both studies were approved by the Human Research Ethics Committee of the University of the Witwatersrand (Medical) and the institutional review board of Boston University Medical Campus. All SLATE study participants provided written informed consent.

I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.

Yes

I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).

Yes

I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.

Yes

All data results produced in the present study are contained in the manuscript and supplementary material. Source data for the SLATE model are available online at Boston University's data repository. Source data for the PREDICT models are owned by the South African Government and were used under license for the current study. Access to these is subject to restrictions owing to privacy and ethics policies set by the South African Government, so they are not publicly available. Requests to access these should be directed to pedro.pisa@righttocare.org.

https://open.bu.edu/handle/2144/44321