Poster presentation at:
9th IAS Conference on HIV Science
23rd July 2017 - 26th July 2017
Revell AD1, Wang D1, Hamers R2, Morrow C3, Wood R3, Reiss P2, 4, van Sighem A4, Johnson M5, Ruiz L6, Alvarez-Uria G7, Sierra-Madero J8, Montaner J9, Lane HC10, Larder BA1 on behalf of the RDI study group.
1 The HIV Resistance Response Database Initiative (RDI), London UK
2 Dept of Global Health, Academic Medical Centre of the University of Amsterdam, Amsterdam Institute for Global Health and Development, Amsterdam, The Netherlands
3 Desmond Tutu HIV Centre, Cape Town, South Africa
4 Stichting HIV Monitoring, Amsterdam, The Netherlands
5 Royal Free Hospital, London, UK
6 Fundacion IrsiCaixa, Badelona, Spain
7 Rural Development Trust (RDT) Hospital, Bathalapalli, India
8 Instituto Nacional de Ciencias Medicas y Nutricion SZ, Mexico
9 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
10 National Institute of Allergy and Infectious Diseases, Bethesda, MD, USA
Optimisation of antiretroviral therapy in resource-limited settings (RLS) can be challenging without access to the newest drugs, or genotyping to help tailor therapy for the individual. Computational models have been developed to predict virological response to therapy but many require a genotype and have been developed using western data. Here we describe the development of new models that do not require a genotype, using a large data set that includes substantial data from RLS and including, for the first time, tipranavir, maraviroc and elvitegravir.
Random forest (RF) models were trained to predict the probability of virological response (viral load <50 copies/ml) to a new regimen following virological failure, using baseline viral load, CD4 count, treatment history, new drug regimen and time to follow-up from 52,270 treatment change episodes (TCEs), including 5,329 from RLS (4,190 from South Africa). The models were assessed during cross-validation, with an independent test set (n=3,000 including 461 from South Africa). The area under the ROC curve (AUC) was the main outcome. The predictive accuracy of the models was compared to genotyping using three rules-based interpretation systems to derive genotypic sensitivity scores (GSS).
The models achieved a mean AUC of 0.83 during cross validation and 0.81 with the 3,000 TCE test set. Sensitivity was 73%, specificity 76% and overall accuracy 75%. For the South African test cases the models achieved an AUC of 0.76, sensitivity of 72%, specificity of 69% and overall accuracy of 70%. The AUC values for cases involving tipranavir, maraviroc and elvitegravir, ranged from 0.75 to 0.89. Of the 3,000 test TCEs, 634 had genotypes and the AUC for the GSS ranged between 0.55 and 0.57 - significantly inferior to the accuracy of the models (p<0.0001).
These new models, trained with our largest dataset so far, were able to predict response to HIV therapy significantly more accurately than genotyping with rules-based interpretation. They were accurate for cases from South Africa and cases involving new drugs. The models are freely available online at https://www.hivrdi.org/treps and have the potential to predict and avoid treatment failure by identifying potentially effective, alternative regimens.