641    

Background:  Therapeutic drug monitoring allows for individual pharmacologic evaluation, but is limited by the interpretation of the individual concentrations. The objective of this study was to develop a computer-based expert system for modeling and interpreting pharmacokinetic data for lopinavir (LPV) and efavirenz (EFV).

Methods:  Data were extracted from CCTG 578, a prospective study of therapeutic drug monitoring in 199 antiretroviral (ARV)-naïve and -experienced patients. An expert committee of HIV clinicians and pharmacologists evaluated real-time pharmacokinetic data and recommended therapy to study investigators. From these recommendations, decision algorithms were generated to interpret plasma concentrations. These algorithms formed the knowledge base of the expert artificial intelligence system. The expert committee modeled LPV and EFV concentrations using a Bayesean nonlinear curve fitting approach to estimate C4 and C_trough. The artificial intelligence system modeler used a set of polynomial equations generated from Monte Carlo simulations to compute estimated trough and 4-hour post-dose pharmacokinetic metrics. The artificial intelligence system then integrated clinical data, including; prior ARV history, HIV RNA treatment response, CD4 cell counts, and HIV phenotypic drug susceptibility to generate a therapeutic drug monitoring recommendation. Paired t-tests were used to compare numerical means of estimated trough and 4-hr post dose concentrations between those generated by the expert committee and the artificial intelligence system. κ statistic was used to measure the inter-rater agreement between the expert committee recommendations to make dose adjustments or not with the artificial intelligence recommendations.

Results:  This analysis included 66 patients on LPV, 47 on EFV, and 3 on both drugs. Correlations were high for both LPV and EFV in 4-hour and 12-hour predicted concentration between the expert committee and artificial intelligence systems (r > 0.89 for all comparisons, p < 0.0001).There was no statistical difference between artificial intelligence and expert committee therapeutic drug monitoring modeled values for trough concentrations of either drug, but differences were seen for predicted 4-hour concentrations:  for EFV the mean concentrations were, 4.2 vs 3.9 (artificial intelligence vs expert committee, p = 0.02), and for LPV the values were 7.9 vs 8.6 (artificial intelligence vs expert committee, p < 0.001). Agreement between expert committee and artificial intelligence therapeutic drug monitoring recommendations were seen in 53 of 69 LPV cases (κ = 0.53, p < 0.001) and 47 of 49 EFV cases (κ = 0.91, p < 0.001).

Conclusions:  Good agreement was seen between the expert committee and artificial intelligence modeled values for trough concentrations. The artificial intelligence system successfully predicted committee recommendations for EFV better than LPV, though both had good κ statistics.

Keywords: AI System; drug concentration; Therapeutic Drug Monitoring