Computational models developed without a genotype for resource-poor countries predict response to HIV treatment with 82% accuracy

Oral and poster presentation at:

XVIII International HIV Drug Resistance Workshop

9th June 2009 - 13th June 2009

Fort Myers, USA


Revell AD 1, Wang D 1, Harrigan R 2, Gatell J 3, Ruiz L 4, Emery S 5, Perez-Elias MJ 6, Torti C 7, Baxter J 8, DeWolf F 9, Gazzard B 10, Geretti AM 11, Staszewski S 12, Hamers R 13, Wensing AMJ 14, Lange J 15, Montaner JM 2, Larder BA 1.

1 HIV Resistance Response Database Initiative (RDI), London, UK
2 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
3 Hospital Clinic of Barcelona, Barcelona, Spain
4 FundaciA3 irsiCaixa, Badalona, Spain
5 National Centre in HIV Epidemiology and Clinical Research, Sydney, Australia
6 RamA3n y Cajal Hospital, Madrid, Spain
7 University of Brescia, Brescia, Italy
8 Cooper University Hospital, NJ, Camden, USA
9 Netherlands HIV Monitoring Foundation, Amsterdam, The Netherlands
10 Chelsea and Westminster Hospital, London, UK
11 Royal Free Hospital, London, UK
12 Hospital of the Johann Wolfgang Goethe-University, Frankfurt, Germany
13 PharmAccess Foundation, Academic Medical Centre, Amsterdam, The Netherlands
14 University Medical Centre, Utrecht, The Netherlands
15 Academic Medical Centre, Amsterdam, The Netherlands.

Background
Optimising therapy following initial treatment failure in resource-poor countries is challenging due to limited treatment options and the relative unavailability of resistance testing. We have demonstrated that computational models trained with clinical data can predict virological response from genotype, viral load, CD4 count and treatment history and may be a useful treatment-selection tool. Here we model response without genotype for potential use in resource-poor countries.

Methods
Two Random Forest models were trained to predict the probability of the follow-up viral load being <50 copies/ml after a treatment change.
Model 1 was trained with 8,214 treatment change episodes (TCEs), representative of clinical practice in resource-poor countries (no previous protease inhibitors (PIs), T-20, raltegravir or maraviroc, with PIs allowed in the new regimen). The input variables were baseline viral load, CD4 count, treatment history (previous AZT, 3TC, any NNRTI), the drugs in the new regimen and time to follow-up.
Model 2 was as Model 1 except 11 individual drug treatment history variables were used.

Receiver operator characteristic curves were plotted for the modelsaEUR(TM) performance with an independent test set of 400 TCEs, from different patients to those used for training. The models were used to identify alternative 3-drug regimens available in resource-poor countries that were predicted to reduce the viral load to <50 copies for the 205 test cases with incomplete response.

Results
Both models achieved an accuracy of 82% and an area under the curve of 0.88. Sensitivity was 77% and 79% and specificity 86% and 85% for Models 1 and 2 respectively. The optimal operating points were 0.44 and 0.47. The most important variable in the modelling was baseline viral load. The two models identified alternative regimens that were predicted to be fully suppressive for 94 (46%) and 98 (48%) of the failures. When all currently licensed protease inhibitors were included, this rose to 100 (49%) and 106 (52%).

Conclusions
Models trained with sufficiently large, representative datasets can predict virological response accurately without a genotype. The results highlight the value of viral load information. This approach has potential for optimising antiretroviral therapy in resource-poor countries.

Slides in PDF View Slides
RDI

Join the RDI mailing list

Enter your email address below to start receiving the RDI newsletter.

You can unsubscribe at any time. View the RDI privacy policy.

Email