Personalized drug concentration predictions with machine learning: an exploratory study

Abstract

Background: The dose individualization by therapeutic drug monitoring (TDM) can be improved if population-based reference ranges are available, as there is large inter- and intrapatient variability. If these ranges are not available, dose individualization may not be optimal. Machine learning can help achieve accurate drug dose settings and predict the resultant levels.Methods: Two random forest models, a multi-class classifier to predict dose and a regression model to predict blood drug level were trained on 320 patients’ data, consisting of their age, sex, dose and blood drug level. The classifier consisted of 1000 estimators (decision trees) and the regression model consisted of 1300 estimators. The model was evaluated on randomly split test set having 10% of the total dataset size. The regression model was compared against k-Nearest neighbor and linear regression models. The classifier was evaluated using accuracy, precision, and F1 Score; the regression model was evaluated using R2, Root mean squared error, and mean absolute error.Results: The classifier had an out-of-sample accuracy of 68.75%, average precision of 0.7567, and an average F1 score of 0.6907. The regression model had an out-of-sample R2 value of 0.2183, root mean squared value of 3.7359, and a mean absolute error of 2.5156. These values signify an average classification performance, and a below-average regression performance due to small dataset.Conclusions: It is possible for machine learning algorithms to be used in therapeutic drug monitoring. With a well-structured, rich, and large dataset, a very accurate model can be built.