Poster presentation at:
International HIV & Hepatitis Drug Resistance Workshop
8th June 2010 - 12th June 2010
Larder BA1, Wang D1, Revell AD1, Emery S2, DeWolf F3, Nelson M4, Perez-Elias MJ5, Harrigan PR6, Montaner JS6.
1 The HIV Resistance Response Database Initiative (RDI), London, UK
2 NCHECR, Sydney, Australia
3 Netherlands HIV Monitoring Foundation, Amsterdam, The Netherlands
4 Chelsea and Westminster Hospital, London, UK
5 Ramon Y Cajal Hospital, Madrid, Spain
6 BC Centre for Excellence in HIV/AIDS, Vancouver, BC, Canada
Background Individualizing antiretroviral drug therapy to overcome/prevent viral drug resistance and maximise response can be challenging. Genotyping with rules-based interpretation is routinely used to support such decisions. Here we describe the development and testing of novel mathematical models designed to predict the probability of virological response as the potential engine for a treatment selection tool.
Methods Random Forest (RF) models (n=10) were developed, using data from clinical cohorts in several countries, to predict the probability of a follow-up viral load <50 copies/ml after a change to antiretroviral treatment (ART). 5,752 treatment change episodes (TCEs) were used comprising baseline viral load, CD4 count and genotype (collected while on failing therapy), treatment history, the drugs in the new regimen and time to follow-up. The accuracy of the models in predicting the virological responses observed in the TCEs was evaluated during 10 x cross validation by plotting ROC curves for the individual models and for their combined outputs. A second evaluation was performed using an independent test set of 50 TCEs from a single clinic. The models were compared with genotypic sensitivity scores (GSS) derived using three interpretation systems in common use (Stanford HIVDB, REGA V8.0.2 and ANRS V2009.07, accessed 16/03/10) in terms of the accuracy of the their predictions for the 50 test TCEs.
Results The models performed with an overall accuracy of 77% (72-81%) and an area under the ROC curve of 0.82 (0.77-0.87). Sensitivity was 67% (62-80%) and specificity was 81% (75-89%). When tested with the independent test set, the models performed with an overall accuracy of 76% (71-80%), the AUC was 0.83 (0.70-0.88). Sensitivity was 71% (65-81%) and specificity 82% (76-87%). The RF models were significantly more accurate as predictors of response than the GSS for the test TCEs (mean overall accuracy of 54% and an AUC of 0.51, p<0.01).
Conclusions The RF models were accurate predictors of virological response to ART. The models were more accurate in their predictions than were GSS derived from genotype interpretation systems in common use as an aid to treatment decision-making. The combination of RF models is being taken forward to power an experimental treatment-decision support tool.