O116. The EuResist expert model for customised HAART optimisation: 2010 update and extension to newest compounds

Background

The design of an optimal highly active antiretroviral therapy (HAART) customised on patient's background and viral genotyping, is still a challenge. EuResist has been the first data-driven system to be implemented as a free web-service for customised HAART optimisation, and was proven to be superior to all existing genotypic interpretation systems (GIS) since it takes into account not only genotype but multiple other variables.

Data and methods

The EuResist database stores and updates periodically demographic, clinical, and genomic information of HIV+ patients from several countries in Western Europe. The EuResist system is trained on treatment change episodes (TCE) drawn from the EuResist data base, composed of a new drug regimen with a baseline HIV-1 RNA load and a CD4+ count, a baseline viral pol genotype, demographic and previous treatment information. Each TCE is associated to an HIV-1 RNA measurement after 8 weeks, which is used for the definition of virological success (below 500 copies/ml or >2 Log10 reduction from baseline HIV-1 RNA). The system is a combination of three independent machine learning models (based on logistic regression, random forests, and Bayesian networks). The 2010 update has been trained on >5,000 TCE, composed of 20 FDA/EMEA approved nucleoside/tide, non-nucleoside, and protease inhibitors, including the recently approved compounds (RAC) tipranavir, etravirine and darunavir.

Results

The EuResist combined system performance in predicting the correct virological success of a TCE after 8-weeks (validation set, n=561) exhibited an area under the receiver operating characteristic (AUROC) of 0.8 (whereas Stanford HIVdb GIS assessed to 0.73, p=0.002). The inclusion of therapy history, clinical, and demographic covariates was shown to increase significantly prediction performance. In the subset of regimens containing RAC (n=151), the EuResist AUROC was 0.7 and Stanford HIVdb AUROC was 0.63, not allowing to assess a significant difference owing to the small sample size. See Figure 1.

Figure 1

Conclusions

Based on patient’s information and virus genotype, the EuResist web-service ranks the most effective HAART regimens by the probability to achieve an undetectable HIV-1 RNA load after 8 weeks. Thus, it might be useful in clinical practice. The 2010 update includes also RAC and achieved fair performance, which is expected to increase with expanding training data.