Poster presentation at:
27th February 2011 - 2nd March 2011
Ene L1, Duiculescu D1, Revell AD2, Wang D2, Youle M3, Pozniak A4, Montaner J5 and Larder BA2.
1 "Dr. Victor Babes" Hospital for Infectious and Tropical Diseases, Bucharest, Romania;
2 RDI, London, UK;
3 HIV Training and Resource Initiative, London, UK;
4 Chelsea and Westminster Hospital, London, UK;
5 BC Centre for Excellence in HIV and AIDS, Vancouver, Canada
In Romania we have a cohort of HIV-infected young adults (clade F) with a decade of treatment experience for whom optimising therapy now presents a major challenge. The RDI has developed computational models that predict virological response to therapy as part of an on-line treatment selection tool. However, these require a genotype (not routinely available in Romania) and have been most accurate for clinics and settings that contributed training data. Here we test a novel model that does not require a genotype, but which was trained without Romanian data, with a test dataset from our cohort.
A random forest model was developed to predict the probability of the HIV viral load (VL) being reduced to <50 copies/ml following the introduction of a new cART. The input variables were baseline VL, CD4 count, treatment history and time to follow-up. The model was developed mainly on subtype B, with 3,188 treatment changes episodes (TCEs) from North America, Western Europe and Australia. The accuracy of the model during cross validation was 78% (AUC = 0.86).
The model made predictions for 39 TCEs, and these were compared to the observed VLs at follow-up. VL monitoring of some cases was performed with assays with limits of detection higher than 50 copies. Three analyses were therefore performed: 1. All cases with a VL recorded as <400 were treated as being <50 and included in the analysis; 2. Values of <400 were omitted from the analysis but cases with a cut-off between 50 and 400 (e.g. <80) were included as <50 (n=30); 3. All cases with a VL cut-off above 50 were omitted (n=25).
The TCEs were from pre-treated patients with a median baseline VL of 3.62 log and CD4 count of 369. In analysis 1, the overall accuracy of the model was 67% and the AUC was 0.60. In analysis 2, the accuracy was 70% and AUC was 0.74 and in analysis 3 the accuracy was 79% and AUC was 0.83. Four of the nine TCEs removed for analysis 3 that had follow-up viral loads recorded as <400 were predicted by the models in analysis 1 to be failures (VL ?50 copies).
The model predicted treatment responses in this group of pre-treated Romanian patients with clade F virus accurately, without the need for a genotype and despite having been trained without data from Romania. The results suggest that computational models might be a useful tool in optimising therapy for treatment-experienced patients in countries with limited resources where genotyping is not always available.