Poster presentation at:
49th Infectious Diseases Medical Congress, Microbiology Medical Congress (ICAAC)
12th September 2009 - 15th September 2009
San Francisco, USA
Larder BA1, Wang D1, Revell AD1, Harrigan R2, Gatell J3, Ruiz L4, Emery S5; Torti C6, DeWolf F7, Pozniak A8, Montaner JM2.
1 RDI, London, UK
2 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
3 Hospital of Barcelona, Spain
4 FundaciA3 irsiCaixa, Badalona, Spain
5 NCHECR, Sydney, Australia
6 University of Brescia, Brescia, Italy
7 Netherlands HIV Monitoring Foundation, Amsterdam, The Netherlands
8 Chelsea and Westminster Hospital, London, UK.
Background
Computational models can predict response to HIV therapy from data including genotype (GT) and show potential as therapy selection tools following virological failure. We compare the accuracy of models with and without GT to assess the potential of the latter as a clinical tool in resource-poor settings.
Methods
Two Random Forest (RF) models were trained to predict the probability of the follow-up viral load being <50 copies/ml after a therapy change using data from 3,188 treatment change episodes (TCEs) including baseline viral load and CD4 count, treatment history and time to follow-up. Model 1 was also trained with the baseline GT. Receiver operator characteristic curves were plotted for the modelsaEUR(TM) performance with an independent test set of 100 TCEs. Genotypic Sensitivity Scores (GSS) were calculated for the test set using the Rega, ANRS and Stanford HIVdb rules-based interpretation systems and used as a predictor of virologic response.
Results
RF Model 1 achieved accuracy of 82%, an area under the curve (AUC) of 0.88, sensitivity of 76% and specificity of 86%. Model 2 achieved accuracy of 78%, AUC of 0.86, sensitivity of 89% and specificity of 71%. The most important variable in the modelling was baseline viral load. The accuracy of the ANRS, REGA and Stanford HIVdb was 66%, 63% and 67% respectively, with AUC of 0.72, 0.68 and 0.71. Both RF models performed significantly better than each rules-based system (P<0.01).
Conclusions
RF models without GT predict virologic response only marginally less accurately than models with GT and both were superior to rules systems. This approach has potential clinical application in resource-poor settings.