Poster presentation at:
14 th International HIV Drug Resistance Workshop.
Quebec, Canada; 7-11 June 2005

Treatment History and Adherence Information Significantly Improves Prediction of Virological Response by Neural Networks

B.A. Larder 1 , D. Wang 1 , A. Revell 1 , R. Harrigan 2 , J. Montaner 2 , S. Wegner 3 , and C. Lane 4 .

1: The HIV Resistance Response Database Initiative (RDI), London, UK; 2: The BC Centre for Excellence in HIV/AIDS Vancouver, Canada; 3: US Military HIV Research Program, Rockville, USA; 4: National Institute of Allergy and Infectious Diseases (NIAID), Bethesda, USA.

Objectives: Inclusion of treatment history data as a 'surrogate' for minority mutant populations in Artificial Neural Network (ANN) models may enhance prediction of virological response to combination therapy. Additionally, use of 'training' data only from highly-adherent patients could also increase the accuracy of ANN models. Four studies were conducted to address these hypotheses.

Methods:
Four sets of ANN models were trained using baseline viral load, drugs in the treatment regimen, resistance mutations (55 in the main models) and time to follow-up as input variables:

1. With and without additional treatment history input variables ( AZT, 3TC, any NNRTI and any PI) (2,559 treatment change episodes (TCEs) in training set)
2. Using data from patients with low and high (< or 90%) adherence (estimated by prescription re-fills, n=623 TCEs)
3. Using data from adherent patients, plus the additional treatment history variables (n=495 TCEs)
4. Using data from adherent patients, a reduced set of 36 mutation input variables (those with greatest impact on virological response), plus the additional treatment history variables (n=495 TCEs).

In each case the ANN models were tested by being given the input variables from an independent dataset, from which they predicted the virological response.

Results: Correlations ( ) between predicted and actual viral load changes were:

1. Basic and treatment history models: 0.30 and 0.45 respectively (p<0.01)
2. Low and high adherence models: 0.11 and 0.29 respectively (p<0.01)
3. Combined high adherence with treatment history models: 0.24
4. Combined high adherence with treatment history models with the reduced mutation set: 0.46.

Conclusions: Treatment history data significantly improved the accuracy of ANN in predicting virological response possibly acting as a surrogate for minority mutant populations. ANN models trained using data from highly adherent patients were significantly more accurate than those trained with data from less adherent patients. Initial combined 'treatment history/adherent' models achieved an intermediate level of performance, possibly due to the relatively small training set and large number of input variables. Combined models trained using a reduced mutation set showed improved performance.