Tacrolimus is one of the most widely used calcineurin inhibitors for prevention of organ rejection after transplantation. It is included in numerous therapeutic regimens; however, it's use is limited by its narrow therapeutic window, which calls for therapeutic drug monitoring. Insufficient plasma concentrations can lead to organ rejection, whereas excessive concentrations lead to an increase in adverse effects, such as nephrotoxicity. Tacrolimus has high intra- and interindividual pharmacokinetic variability, which is influenced by numerous genetic, clinical and demographic factors, along with drug-drug interactions.
In the context of this master’s thesis, we aimed to evaluate the possibility of using machine-learning methods for predicting tacrolimus whole blood concentration in patients post lung transplantation, which could improve the dosing regimens design. In collaboration with University Medical Centre Ljubljana we performed a retrospective study, which included 54 patients post lung transplantation. Our analysis included tacrolimus concentrations in whole blood and corresponding doses, demographic factors, comedication, biochemical parameters and genotyping results.
For tacrolimus concentration prediction, we used publicly accessible machine learning packages in open-source software environment R, and compared results between an artificial neural network model and gradient-boosted trees (XGBoost) on two differently prepared data sets. We achieved better results using XGBoost, which were R2 0,52 , RMSE 2,48, MAE 1,63 in MAPE 18,57 %, 67,03 % of predicted values were within 20 % deviation from measured whole blood tacrolimus concentrations. Going forward, an individualized software programme and optimised data collection would be needed to improve our data analysis and prediction accuracy. Both developed models were static, and because our data is dynamic as we followed the patients over a longer time period, machine learning methods appropriate for time series analysis might yeild better results.
|