Poster presentation at:
International AIDS Society Conference (IAS)
17th July 2011 - 20th July 2011
Larder BA1, Revell AD1, Wang D1, Ene L2, Tempelman H3, Barth RE4, Wensing AM4, Gazzard B5, DeWolf F6, Lane HC7, Montaner JSS8.
1 RDI, London UK
2 "Dr. Victor Babes" Hospital for Infectious and Tropical Diseases, Bucharest, Romania
3 Ndlovu Care Group, Elandsdoorn, South Africa
4 UMC Utrecht, Netherlands
5 Chelsea and Westminster Hospital, London, UK
6 Netherlands HIV Monitoring Foundation, Amsterdam, The Netherlands
7 NIAID, Bethesda, USA
8 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
The RDI has pioneered the development of computational models that predict response to antiretroviral therapy from baseline data including genotype, typically with 80% accuracy, which are now available online. Genotyping is not generally affordable in resource-limited settings and here we describe the development of models that do not require a genotype, for use in such settings.
Treatment change episodes (TCEs) from >15 countries from the RDI database were partitioned at random into a training set of 14,964 and a test set of 800. A committee of 10 random forest models was developed to predict the probability of follow-up viral load ?400 copies using 10x leave-n out cross-validation (CV). The input variables were baseline viral load, CD4 count, treatment history, drugs in the new regimen and time to follow-up. Receiver-operator characteristic (ROC) curves were plotted during cross validation, with the 800 test set and with test sets from Romania (n=39) and South Africa (n=56). A subset of 57 TCEs from the 800 test set that had genotypes available were also tested with the RDI's models that include a genotype.
The mean area under the curve (AUC) and overall accuracy (OA) were 0.77 and 72% during CV and 0.77 and 71% with the test dataset. With the Romanian test set the AUC and OA were 0.68 and 67% and with the South African test set 0.69 and 68%. The RDI genotype models achieved an AUC and OA of 0.77 and 74% with the 57 subset, compared to 0.76 and 68% for the 'no-genotype' models.
The RF models' predictions were only slightly less accurate than models that include the genotype. They performed well with data from resource-limited clinics not represented in the training dataset suggesting they are generalisable. The models are now available via the RDI's online treatment selection tool HIV-TRePS.