Poster presentation at:
XIV International HIV Drug Resistance Workshop
7th June 2005 - 11th June 2005
Quebec, Canada
Wegner S1 , Larder BA2 , Wang D2 , Revell AD2 , Harrigan R3 , Montaner J3 and Lane C4.
1: US Military HIV Research Program, Rockville, MD, USA; 2: RDI, London, UK; 3: The BC Centre for Excellence in HIV/AIDS Vancouver, BC, Canada; 4: National Institute of Allergy and Infectious Diseases (NIAID), Bethesda, MD, USA.
Background
Previous RDI studies suggested that 'local' Artificial Neural Network (ANN) models trained with data from a single clinic may more accurately predict virological response to combination therapy from genotype than models trained with additional data from two to three other centres. Here we compare local models with 'global' models trained with data from over 200 clinics.
Methods
40 and 38 treatment change episodes (TCEs) from clinics A and B, were partitioned from the RDI database to act as test data sets. Ten local ANN models were trained with the remaining data from clinic A (337 TCEs) and ten global models with 3,168 TCEs from >200 clinics, including the 337 from clinic A. The models were then tested with the two test sets: the models predicted viral load (VL) response from baseline VL, resistance mutations and drug regimen; the predictions of all 10 models were averaged for each TCE and the 'committee' prediction compared with the actual VL from each test set. The committees were then used to identify alternative regimens for cases of treatment failure from clinic A.
Results
Test set A: Correlations between predicted and actual VL gave values of 0.70 vs 0.78 for global and local models respectively and the mean absolute difference between predicted and actual VL was 0.55 vs 0.49 logs (ns). Test set B: Correlations between predicted and actual VL gave values of 0.23 vs 0.07 and the mean differences between predicted and actual VL were 0.66 vs 0.97 logs (p<0.05) for the global and local models respectively.
Both sets of models identified effective alternative regimens for the 19 treatment failures. The mean reduction in VL predicted for the best alternative was 1.83 logs and 1.71 logs for the global and local committees respectively (ns).
Conclusions
Global ANN models can perform as accurately as local models, trained with data from a single clinic, in predicting response for patients from that clinic. Global models appear superior to local models for other clinics, suggesting that they may be the most powerful way to exploit ANN as a generally available treatment decision-making tool.
