Poster presentation at:
XVII International HIV Drug Resistance Workshop
10th June 2008 - 14th June 2008
Sitges, Spain
Larder BA1, Wang D1, Revell AD1, Coe D1, Torti C2, Montaner JSG3, Harris M3, Lane HC4.
1 HIV Resistance Response Database Initiative (RDI), London, UK
2 University of Brescia, Brescia, Italy
3 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
4 National Institute of Allergy and Infectious Diseases (NIAID), Bethesda, MD, USA.
Background
Virologic response to combinations of antiretroviral drugs is highly complex and difficult to predict. We have demonstrated that computational models trained with substantial data from clinical practice can predict response from genotype, viral load, CD4 count and treatment history. Here we present interim results from an ongoing prospective pilot study of our system in clinical practice.
Method
A random forest model and a committee of 10 artificial neural network models were trained to predict virological response (DeltaVL) using data from 3,188 treatment change episodes (TCEs) from multiple clinical sources. These models were tested using an independent set of 100 TCEs and their predictions correlated with the actual DeltaVL values with an RSquared value of 0.68 and a mean absolute difference (predicted vs actual DeltaVL) of 0.49 log copies/ml. An on-line treatment decision tool was developed through which physicians enter the data for a patient requiring a treatment change and receive a report listing the five combinations of drugs the models predict will result in the largest DeltaVL. The physicians were then required to enter their final treatment decision and a follow-up viral load at 12 weeks.
Results
Eight patients had completed the study at the time of this analysis. The correlation (RSquared) between 12-week DeltaVL and the modelsaEUR(TM) predictions was 0.58 and the absolute difference score was 0.47 log copies/ml. In five cases the physician changed their original treatment intention after receiving the RDI report. Physicians indicated that the system was easy to use and would be utilised aEUR~quite oftenaEUR(TM) if generally available.
Conclusions
These preliminary results indicate that the models are able to predict virological response to HAART with accuracy in clinical practice, that the system affects treatment decision-making and would be widely used if made available. The clinical pilot study continues and an international, multi-centre prospective controlled clinical trial of the RDI system is planned.
Acknowledgement
Funded by NCI Contract N01-CO-12400
