Poster presentation at:
International Workshop on HIV & Hepatitis Virus Drug Resistance and Curative Strategies
5th June 2012 - 9th June 2012
Revell AD1, Wang D1, Streinu-Cercel A2, Ene L3, De Wolf F4, Gazzard B5, Gatell J6, Ruiz L7, Perez-Elias MJ8, Montaner JSG9, Lane HC10, Larder BA1 on behalf of the global RDI study group.
1 The HIV Resistance Response Database Initiative (RDI), London, UK.
2 The National Institute of Infectious Diseases "Prof.Dr. Matei Bal?", Bucharest, Romania
3 "Dr. Victor Babes" Hospital for Infectious and Tropical Diseases, Bucharest, Romania
4 Netherlands HIV Monitoring Foundation, Amsterdam, The Netherlands
5 Chelsea and Westminster Hospital, London, UK
6 Hospital Clinic of Barcelona, Barcelona, Spain
7 Fundacio irsiCaixa, Badalona, Spain
8 Ramon y Cajal Hospital, Madrid, Spain
9 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
10 NIAID, Bethesda, USA
Background The optimal selection of antiretroviral drugs following treatment failure is challenging, e.g. due to the complexity of resistance and range of treatment options in some settings or the unaffordability of genotyping and some drugs in others. Computer models have been developed that accurately predict response to therapy to aid decision-making. However, their performance may not be generalisable to data unrepresented in the training set. Here we test recent models with independent data from unfamiliar settings.
Methods Two sets of random forest models, developed using data from USA, Canada, Europe and Australia, were evaluated. All predicted the probability of response from a range of variables, including baseline viral load and CD4 count, ?18 treatment history variables, time to follow-up and the drugs in the new regimen. The models' unique characteristics were:
Set 1: Do not require a genotype, predict follow-up viral load ?400 copies, trained using 14,964 treatment change episodes (TCEs) and achieved an AUC of 0.77 and 72% accuracy during development.
Set 2: Require a genotype (62 mutations) predict follow-up viral load ?50 copies, trained using 7,263 TCEs and achieved an AUC of 0.82 and 76% accuracy during development.
The generalisability of Set 1 was assessed using 375 TCEs without genotypes from Romania and Set 2 using 375 TCEs with genotypes from clinics in North America and Europe not represented in the training data.
Results Set 1 (no genotype) achieved a mean AUC of 0.73 and accuracy of 67%. Set 2 achieved a mean AUC value of 0.87 and accuracy of 91%. Mean sensitivity and specificity values were 65% and 70% for Set 1 and 62% and 92% for Set 2. The models identified alternative regimens that were predicted to be effective in 97% and 98% of cases where the new regimen used in the clinic had failed.
Conclusions The models were accurate for new data from unfamiliar settings and were able to identify potentially effective regimens following treatment failure. While the models that use a genotype performed significantly better than those that do not, the performance of the 'no-genotype' models with data from an unfamiliar setting (Romania) was encouraging.