Oral and poster presentation at:
International Workshop on HIV & Hepatitis Drug Resistance and Curative Strategies
7th June 2011 - 10th June 2011
Los Cabos, Mexico
Larder BA1, Revell AD1, Wang D1, Hamers R2, Tempelman H3, Barth R4, Wensing AMJ4, Morrow C5, Wood R5, DeWolf F6, Kaiser R7, Pozniak A8, Lane HC9, Montaner JM10.
1 HIV Resistance Response Database Initiative (RDI), London, UK
2 PharmAccess Foundation, Academic Medical Centre, Amsterdam, The Netherlands
3 Ndlovu Care Group, Elandsdoorn, South Africa
4 University Medical Centre, Utrecht, The Netherlands
5 Desmond Tutu HIV Centre, Cape Town, South Africa
6 Netherlands HIV Monitoring Foundation, Amsterdam, The Netherlands
7 AREVIR cohort, University of Cologne, Cologne, Germany
8 Chelsea and Westminster Hospital, London, UK
9 National Institute of Allergy and Infectious Diseases, Bethesda, MD, USA
10 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
Background
Optimising therapy following treatment failure in resource-limited settings (RLS) is challenging due to limited treatment options and the general unavailability of resistance testing. We have demonstrated that computational models can predict virological response from genotype, viral load, CD4 count and treatment history and are a useful web-based clinical tool. Here we develop models that do not require a genotype and, for the first time, assess their potential clinical utility using data from RLS.
Methods
10 random forest models were trained to predict the probability of a follow-up viral load being <400 copies/ml. The input variables were baseline viral load, CD4 count, treatment history, drugs in the new regimen and time to follow-up. 14,891 treatment change episodes (TCEs), predominantly from Europe and North America, were used for training with 800 set aside at random as an independent test set. Additional test sets were provided by two South African clinics, Gugulethu (114 TCEs) and Elandsdoorn (39 TCEs) and the PASER-M cohort in six sub-Saharan African countries (78 TCEs). Relevant TCEs were extracted from PASER and added to Elandsdoorn and Gugulethu to give a South African test set. Model performance was assessed using ROC curves, in terms of area under the curve (AUC) and overall accuracy.
Results
The models achieved a mean AUC of 0.77 (95%CI 0.76,0.78) and accuracy of 75% during cross validation and 0.77 (95%CI 0.73,0.80) and 71% with the 800 test set. For the three RLS test sets, AUC ranged from 0.58-0.65 and accuracy from 67-72%. For South African TCEs, AUC= 0.62 (95%CI 0.53,0.71) and accuracy = 65%. The optimum operating point was adjusted for the South African data and the models correctly predicted 60% of actual failures and identified alternative 3-drug regimens, using locally available drugs, which were predicted to be effective for 79% of these.
Conclusions
Computational models trained with large datasets from resource-rich countries can predict virological response without a genotype with an accuracy that is comparable to that of genotyping with rules-based interpretation. The models have the potential to predict and avoid treatment failure by identifying effective, alternative, practical regimens. This approach has potential clinical utility for managing cases of treatment failure in RLS.
View Slides