Poster presentation at:
15th International Workshop on HIV Observational Databases
24th March 2011 - 26th March 2011
Prague, Czech Republic
Revell AD1, Wang DW1, Coe D1, Mican, JM2, Agan BK3, Harris M4, Torti C5, Izzo I5, Emery S6, Boyd M6, Ene L7, De Wolf F8, Nelson M9, Metcalf JA2, Montaner JSS4, Lane HC2, Larder BA1.
1 RDI, London, UK
2 NIAID, Bethesda, MD, USA
3 Uniformed Services University of the Health Sciences, Bethesda, MD, USA
4 BC Centre for Excellence in HIV/AIDS, Vancouver, Canada
5 University of Brescia, Brescia, Italy
6 NCHECR, Sydney, Australia
7 "Dr. Victor Babes" Hospital for Infectious and Tropical Diseases, Bucharest, Romania
8 Netherlands HIV Monitoring Foundation, Amsterdam, The Netherlands
9 Chelsea and Westminster Hospital, London, UK
There are 25 antiretroviral drugs available and selection of the optimum combination for cases of virological failure can be challenging given the number of combinations possible, the variety of patients' treatment experience, the complexity of viral resistance and other factors. Here we describe the use of data from multiple cohorts and clinics to train computational models to predict virological response to combination antiretroviral therapy (cART) and assessment of the online treatment support system that uses them.
10 Random Forest (RF) models were developed to predict the probability of a follow-up viral load <50 copies/ml after a change to cART using data from clinics and cohorts in >15 countries. 5,752 treatment change episodes (TCEs) were used comprising baseline viral load, CD4 count and genotype (collected while on failing therapy), treatment history, the drugs in the new regimen and time to follow-up. The accuracy of the models was evaluated during cross validation and also using independent test sets of 50 TCEs from a single clinic in Australia.
An online system was developed through which physicians could enter their patients' data and their initial treatment decision via the RDI website and receive a report listing the RDI models' predictions of response to the physician's selection and to possible alternative regimens. The system was evaluated in two ways: 1. Pre-launch, 114 cases of treatment failure were entered by 23 physicians and they were required to complete an evaluation questionnaire. 2. Post-launch, users were invited to complete the same questionnaire: to date 11 have been submitted.
During cross-validation the models performed with an overall accuracy of 77% (72-81%) and an area under the ROC curve of 0.82 (0.77-0.87). Sensitivity was 67% (62-80%) and specificity was 81% (75-89%). When tested with the independent Australian test set, the models performed with an overall accuracy of 76% and an AUC of 0.83.
Physicians rated entering baseline data as very or quite easy (88%), satisfactory (9%) and quite difficult (3%). The RDI report was rated as very or quite easy to understand (82%) and satisfactory (18%). The system was rated as useful (34%), satisfactory (38%) and not very useful (28%). They indicated they would use RDI very or quite frequently (41%), sometimes (38%), infrequently (15%) and never (6%). Testing of the system pre-launch was associated with a change of treatment decision in one-third of cases and an average saving in the number of drugs in the final regimen of approximately 10%.
Computational models developed using data from multiple cohorts and clinics can make accurate predictions of virological response to antiretroviral therapy for failing patients in different settings. An online treatment support system using the models was rated favourably by physicians who amended their treatment decisions following use of the system in a significant proportion of cases.