Introduction: Real-time electronic adherence monitoring (EAM) systems could inform on-going risk assessment for HIV viraemia and be used to personalize viral load testing schedules. We evaluated the potential of real-time EAM (transferred via cellular signal) and standard EAM (downloaded via USB cable) in rural Uganda to inform individually differentiated viral load testing strategies by applying machine learning approaches.
Methods: We evaluated an observational cohort of persons living with HIV and treated with antiretroviral therapy (ART) who were monitored longitudinally with standard EAM from 2005 to 2011 and real-time EAM from 2011 to 2015. Super learner, an ensemble machine learning method, was used to develop a tool for targeting viral load testing to detect viraemia (>1000 copies/ml) based on clinical (CD4 count, ART regimen), viral load and demographic data, together with EAM-based adherence. Using sample-splitting (cross-validation), we evaluated area under the receiver operating characteristic curve (cvAUC), potential for EAM data to selectively defer viral load tests while minimizing delays in viraemia detection, and performance compared to WHO-recommended testing schedules.
Results: In total, 443 persons (1801 person-years) and 485 persons (930 person-years) contributed to standard and real-time EAM analyses respectively. In the 2011 to 2015 dataset, addition of real-time EAM (cvAUC: 0.88; 95% CI: 0.83, 0.93) significantly improved prediction compared to clinical/demographic data alone (cvAUC: 0.78; 95% CI: 0.72, 0.86; p = 0.03). In the 2005 to 2011 dataset, addition of standard EAM (cvAUC: 0.77; 95% CI: 0.72, 0.81) did not significantly improve prediction compared to clinical/demographic data alone (cvAUC: 0.70; 95% CI: 0.64, 0.76; p = 0.08). A hypothetical testing strategy using real-time EAM to guide deferral of viral load tests would have reduced the number of tests by 32% while detecting 87% of viraemia cases without delay. By comparison, the WHO-recommended testing schedule would have reduced the number of tests by 69%, but resulted in delayed detection of viraemia a mean of 74 days for 84% of individuals with viraemia. Similar rules derived from standard EAM also resulted in potential testing frequency reductions.
Conclusions: Our machine learning approach demonstrates potential for combining EAM data with other clinical measures to develop a selective testing rule that reduces number of viral load tests ordered, while still identifying those at highest risk for viraemia.
Keywords: adherence; machine learning; real-time adherence monitoring; viraemia; viral load monitoring; virologic failure.