Launch of models that cater for missing baseline data and different definitions of virological response
Dear HIV-TRePS user / RDI subscriber
We are writing to you today to inform you of the addition of a range of 12 new classifier models in HIV-TRePS to predict response to HIV drug therapy for cases missing some of the standard baseline data. They also cater for centres that use definition of virological response other than the RDI standard threshold of 50 copies HIV-RNA/ml. All the models achieved a high level of accuracy in independent testing although there was a small loss of accuracy with piece of missing baseline data. All achieved an area under the ROC curve greater than 0.8, and overall accuracy, sensitivity and specificity greater than 70%.
The new models were developed specifically for use in settings with limited resources, where patient monitoring may be incomplete. None of them required a baseline viral genotype and the models were able to perform well despite missing the following additional baseline data:
1. Missing baseline viral load aEUR" estimating the probability of a follow-up viral load <50 copies HIV RNA/mL.
2. Missing baseline viral load aEUR" estimating the probability of a follow-up viral load <400 copies.
3. Missing baseline viral load aEUR" estimating the probability of a follow-up viral load <1,000 copies.
4. Missing baseline time on therapy and viral load aEUR" 50 copies.
5. Missing baseline time on therapy and viral load aEUR" 400 copies.
6. Missing baseline time on therapy and viral load aEUR" 1,000 copies.
7. Missing baseline CD4 count and viral load aEUR" 50 copies.
8. Missing baseline CD4 count and viral load aEUR" 400 copies.
9. Missing baseline CD4 count and viral load aEUR" 1,000 copies.
10. Missing baseline CD4 count, time on therapy and viral load aEUR" 50 copies.
11. Missing baseline CD4 count, time on therapy and viral load aEUR" 400 copies.
12. Missing baseline CD4 count, time on therapy and viral load aEUR" 1,000 copies.
The last three models, missing a baseline genotype, CD4 count, time on therapy and viral load, made their predictions based only on the drugs to which the patient has been exposed in the past and the drugs in the new regimen. They were trained with our largest ever data sets (as we could include cases with more missing data) and their performance was reduced by around 10% relative to models trained using all the baseline data.
All the models were able to identify multiple alternative regimens that were predicted to lead to virological responses for the great majority of cases that failed the new regimen introduced in the clinic.
We believe these latest models represent a major advance with greatly enhanced utility and potential benefit in countries with limited drugs and diagnostics, for example in resource-limited settings.
All of us at the RDI hope that you find these new models, and the HIV-TRePS system as a whole, useful in your local settings.
Date published: 31st August 2021