The development of artificial neural networks to predict virological response to combination HIV therapy.

Methods: As an alternative to current interpretation systems, we have developed artificial neural network (ANN) models to predict virological response to combination therapy from HIV genotype and other clinical information.

Results: ANN models trained with genotype, baseline viral load and time to follow-up viral load (1154 treatment change episodes from multiple clinics), produced predictions of virological response that were highly significantly correlated with actual responses (r2 = 0.53; P 0.00001) using independent test data from clinics that contributed training data. Augmented models, trained with the additional variables of baseline CD4+ T-cell count and four treatment history variables, were more accurate, explaining 69% of the variance in virological response. Models trained with the full input dataset, but only those data involving highly active antiretroviral therapy (three or more full-dose antiretroviral drugs in combination), performed at an intermediate level, explaining 61% of the variance. The augmented models performed less well when tested with data from unfamiliar clinics that had not contributed data to the training dataset, explaining 46% of the variance in response.

Conclusion: ANN models trained with genotype, baseline viral load and time to follow-up viral load (1154 treatment change episodes from multiple clinics), produced predictions of virological response that were highly significantly correlated with actual responses (r2 = 0.53; P 0.00001) using independent test data from clinics that contributed training data. Augmented models, trained with the additional variables of baseline CD4+ T-cell count and four treatment history variables, were more accurate, explaining 69% of the variance in virological response. Models trained with the full input dataset, but only those data involving highly active antiretroviral therapy (three or more full-dose antiretroviral drugs in combination), performed at an intermediate level, explaining 61% of the variance. The augmented models performed less well when tested with data from unfamiliar clinics that had not contributed data to the training dataset, explaining 46% of the variance in response.