CROI 2019 Abstract eBook

Abstract eBook

Poster Abstracts

and 27% (4571/17,042) had a viral load test at each recommended time-point respectively. VL results were documented at all recommended time-points for 11.5% (2613/22,730) and 4.9% (838/17,042) of patients on ART for 12 and 24 months respectively. We documented 12% (2,456/ 20,405) individuals with at least one VL≥1000 copies/mL. Of these, 738 (30%) had a repeat VL within 6 months, and 425 (17%) achieved successful management of virologic failure with either re-suppression or appropriate change to second-line therapy (Figure). For the 150 individuals who switched to second-line, the median time to regimen change was 345 days (IQR 135-671) after their first elevated viral load measurement. Conclusion: We found suboptimal VL monitoring, and delayed or absent responses to VF in public-sector ART clinics in rural South Arica. Such delays are likely to increase the likelihood of patient morbidity, and transmission of drug resistant HIV. We did not investigate howmuch of our finding could be explained by failure to capture VL results in Tier.Net. Future studies should investigate causes of suboptimal VL monitoring and consider what interventions are needed to improve attention to VF in the region.

improved prediction of failure beyond demographic and clinical data alone (c-statistic=0.79; 95% CI: 0.72, 0.87; p=0.05). A hypothetical testing strategy using real-time EAM to decide when to order versus defer viral load testing would have reduced the number of viral load tests by 30%, while still detecting 87% of all virologic failures without additional delay. By comparison, the WHO- recommended testing schedule would have reduced the number of viral load tests by 69%, but resulted in delayed detection of virologic failure a mean of 74 days (SD = 41 days) for >80% of individuals with failure. Conclusion: Our machine learning approach demonstrates potential for combining EAM data with other clinical measures to develop a selective testing rule that may reduce costs incurred by both researchers and patients, while still identifying those at highest risk for virologic failure 1047 PREDICTORS OF ART INITIATION AND VIRAL SUPPRESSION IN A LARGE COHORT IN UKRAINE Kostyantyn Dumchev 1 , Iuliia Novak 2 , Tetiana Saliuk 2 1 Ukrainian Institute on Public Health Policy, Kyiv, Ukraine, 2 Alliance for Public Health, Kyiv, Ukraine Background: Rapid initiation of ART, treatment adherence support, proper management of virologic failure are important strategies for reaching the ambitious 90-90-90 goals in Ukraine and globally. Key national stakeholders and international donors have set ambitious fast track goals to increase the number of patients from 88,270 on 01/01/2018 to 140,000 by the end of 2018. This study was commenced to obtain reliable data on key treatment quality indicators, contributing factors and trends to inform program planning. Methods: Data frommedical charts of all patients who received care at HIV facilities in 2010-2016 in 18 out of 27 regions of Ukraine were entered into an electronic medical record system. After verification of data quality, depersonalized datasets linked by unique patient code were extracted at each facility and merged for analysis. This analysis focused on the effect of clinical variables (HIV mode of transmission, clinical stage, CD4, VL, TB, HCV, injecting drug use [IDU]) on time from diagnosis to ART initiation and to viral suppression (<200cp/ml). The entire dataset, excluding children younger than 15 at diagnosis, was analyzed using Cox proportional hazard models. Results: The cohort included 37,690 patients with HIV infection, approximately 30% of all patients receiving care in Ukraine in 2016. Average age at diagnosis 46.4%were females. Median time from diagnosis to ART was 26 months (95%CI: 25.0-26.9) and 14 months (95%CI: 13.7-14.3) from ART to viral suppression. Multiple significant predictors were identified for both outcomes (see Table). Notably, the time to ART initiation was increasing with male gender (aHR=.91), negative TB status (aHR=.9), being at early clinical HIV stage (aHR=.53), IDU mode of transmission (aHR=.77). The chance of getting ART was increasing with lower CD4 (aHR=4.1 for CD4<200), reporting no recent IDU (aHR=1.11), having positive TB test (aHR=1.18), homosexual mode of transmission (aHR=1.18). Viral suppression was associated with younger age (aHR=.98), earlier clinical stage (aHR=1.08), having negative TB test (aHR=.86), IDU mode of transmission (aHR=.93). Overall, coverage of key clinical assessments was not universal, and completion was associated with both outcomes. Conclusion: Quality of HIV care in Ukraine, characterized by coverage of key clinical tests, time to ART initiation and viral suppression indicators remains suboptimal. Patients with advanced disease had priority for ART, reflecting the delayed adoption of test-and-start strategy.

Poster Abstracts

1046 MACHINE LEARNING APPLIED TO ELECTRONIC ADHERENCE DATA TO INFORM VIRAL LOAD MONITORING Alejandra E. Benitez 1 , Maya L. Petersen 1 , Nicholas Musinguzi 2 , David R. Bangsberg 3 , Yap Boum 4 , Bosco M. Bwana 2 , Conrad Muzoora 2 , Peter W. Hunt 5 , Jeffrey N. Martin 5 , Jessica E. Haberer 3 1 University of California Berkeley, Berkeley, CA, USA, 2 Mbarara University of Science and Technology, Mbarara, Uganda, 3 Massachusetts General Hospital, Boston, MA, USA, 4 Epicentre, Mbarara, Uganda, 5 University of California San Francisco, San Francisco, CA, USA Background: Approaches for tailoring ART monitoring are needed to optimize the impact and cost-effectiveness of differentiated care delivery systems. Real-time electronic adherence monitoring (EAM) could potentially inform ongoing risk assessment for virologic failure, and thus be used to modify viral load testing schedules. We evaluated the potential of EAM data to contribute to an individually differentiated viral load testing strategy by applying machine- learning approaches to real-time EAM data from Uganda. Methods: We evaluated an observational cohort of persons living with HIV who were treated with ART and monitored with EAM (2005-2015). Super Learner, an ensemble machine-learning method, was used to build a risk score for virologic failure (>1000 copies/ml) based on clinical (CD4 count, pre-ART viral load, ART regimen) and demographic data, together with EAM-based adherence. Using sample-splitting (cross-validation), we evaluated the performance of this risk score to determine: 1) whether EAM improved prediction of failure beyond clinical and demographic data; 2) potential for real-time EAM data to selective defer viral load tests while minimizing delays in failure detection; and, 3) performance compared to WHO-recommended testing schedules. Results: 485 individuals (242 of whomwere initiating ART) contributed 2834 outcome viral loads over 930 person-years. Median CD4 at ART initiation was 200 cells/mm3 (IQR 111, 317); 45 patients (1.6%) experienced virologic failure. Super Learning applied to real-time EAM data achieved excellent prediction of virologic failure (cross-validated c-statistic=0.89; 95% CI:0.85, 0.94) and

CROI 2019 411