Poster presentation at:
10th European Meeting on HIV & Hepatitis Treatment Strategies and Antiviral Drug Resistance
28th March 2012 - 30th March 2012
Barcelona, Spain
Revell AD1, Wang D1, DeWolf F2, Gatell J3, Ruiz L4, Nelson5, M, Perez-Elias6, MJ, Lane HC7, Montaner JSS8, Larder, BA1 on behalf of the RDI study group.
1 RDI, London UK
2 Netherlands HIV Monitoring Foundation, Amsterdam, The Netherlands
3 Hospital Clinic of Barcelona, Barcelona, Spain
4 FundaciA3 irsiCaixa, Badalona, Spain
5 Chelsea and Westminster Hospital, London, UK
6 RamA3n y Cajal Hospital, Madrid, Spain
7 NIAID, Bethesda, USA
8 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
Background The optimum selection and sequencing of combination therapy (cART) to maintain viral suppression can be challenging, particularly in settings where access to drugs, information or expertise may be limited. The HIV Resistance Response Database Initiative pioneered the development of computational models that predict the virological response to cART and make these models available on line as the HIV Treatment Response Prediction System (HIV-TRePS). Accurate models have been developed using a range of input variables, including limited treatment history information. Here we describe the development and testing of new random forest models to power HIV-TRePS. We compared models that used limited or comprehensive treatment history information and we included predictions of response to raltegravir for the first time.
Materials & Methods Data for model development were extracted from the RDI database of 83,871 patients. Data filters were applied and 7,263 treatment change episodes (TCEs) were extracted that met all entry criteria with no missing data points. The TCEs were used to train two committees of 5 random forest models to predict the probability of virological response to a new regimen. The input variables were baseline CD4 cell count, viral load and genotype (62 mutations in protease and reverse transcriptase), drugs in the new regimen and time from treatment change to follow-up. Committee 1 also included six binary treatment history variables coding for any previous exposure to: zidovudine, lamivudine/emtracitabine, enfuvirtide, raltegravir. Committee 2 included 18 treatment variables for exposure to individual drugs. The accuracy of the models was assessed during cross-validation by plotting receiveraEUR"operator characteristic (ROC) curves.
Results The models in Committee 1 achieved an area under the curve (AUC) during cross-validation of 0.78aEUR"0.83 (meanaEURS=aEURS0.815), accuracy of 73aEUR"77% (meanaEURS=aEURS75%), sensitivity of 61aEUR"68% (meanaEURS=aEURS64%) and specificity of 80aEUR"84% (meanaEURS=aEURS81%). The models in Committee 2 achieved an AUC of 0.80aEUR"0.84 (meanaEURS=aEURS0.820), accuracy of 75aEUR"78% (meanaEURS=aEURS76%), sensitivity of 60aEUR"65% (meanaEURS=aEURS62%) and specificity of 81aEUR"87% (meanaEURS=aEURS84%). The difference in AUC between the two committees is not statistically significant (p=0.21), using a paired t-test. The AUC values for the 243 TCEs containing raltegravir ranged from 0.66-0.76 (mean=0.71) for Committee 1 and 0.63 to 0.78 (mean=0.71) for Committee 2.
Conclusions The models achieved a consistent, high level of accuracy in predicting treatment responses. The differences between the models using the restricted set of treatment history variables and those using 18 individual drug history variables were minimal. The accuracy of the models for raltegravir TCEs was somewhat reduced compared with the overall performance, reflecting perhaps the relatively small number of raltegravir TCEs and/or the absence of genotype information relating to this inhibitor. The models are being used to power HIV-TRePS, the online aid to treatment selection.