Poster presentation at:
12th International Workshop on HIV Drug Resistance
10th June 2003 - 13th June 2003
Wang D, Larder BA, Revell AD, Harrigan R, Montaner J. On behalf of the HIV Resistance Response Database Initiative (RDI).
Background We have previously demonstrated the utility of neural networks (NN) in
predicting viral load (VL) response using a clinical trial dataset. Here we compare the reliability of using 'real life' clinical cohort data to model treatment response.
Methods 351 patients from the BC HIV treatment cohort were analysed. Each had multiple genotypes and irregular VL measurements. 'Treatment change episodes' (TCEs) were selected from patients having a genotype up to 12 weeks before treatment change and baseline VL up to 16, 12 or 8 weeks before treatment change. Follow-up VL was within 4-40 weeks after treatment change. 652, 602 and 539 TCEs respectively, obtained using the 3 VL windows, were used for NN training (10% of each was partitioned for independent testing). Results were compared to those previously obtained using clinical trial data comprising 700 TCEs, with baseline VL and genotype at Week 0 (time of treatment change), and follow-up VL at 8, 16, or 24 weeks post treatment change.
Results The correlation between the predicted and actual VL change for the 16-, 12- and 8-week baseline VL models gave the following r@ values: 0.42, 0.45 and 0.55 respectively. The model with the narrowest baseline VL window (8 weeks) most accurately predicted VL response. The r@ value for this model was significantly different to those for the other models (p<0.05) and also to that of a previous model derived using clinical trial data (r@ =0.50, p<0.05). The NN models were also used to predict VL trajectory. The rates of correct trajectory prediction were: 74%, 72% and 87% (16-, 12- and 8-week models respectively). Again, the 8-week model was significantly more accurate than the two other models (p<0.05) and comparable to that obtained using the clinical trial model (75%, ns). We utilized the 8-week baseline VL model to evaluate alternative regimens in patients who had experienced treatment failure. The model was used to predict response to 30 common HAART regimens for each of 118 TCEs where VL increased after a change to a regimen of =3 antiretroviral drugs (3,540 evaluations). In 116 cases (98%), one or more regimens were identified that the model predicted would cause a reduction in VL. This could translate into an additional 101 cases where patients would have benefited from treatment selected using this model (based on the 87% accuracy of trajectory prediction established above).
Conclusions NN modeling using the narrowest baseline VL window gave the most accurate prediction of VL response, underlining the importance of the quality of the dataset. This model also showed superior predictive accuracy compared to data obtained from a NN model trained with clinical trial data, establishing the feasibility of using treatment cohort data to model VL response. Furthermore, the model identified potentially beneficial treatment regimens in almost all patients who experienced treatment failure, demonstrating the utility of using this approach to perform evaluations of multiple treatment permutations.