B-141 Assessment of Appropriateness of a Machine Learning-Based Model Compared with the Pharmacokinetic Dosing Systems for Estimating Serum Vancomycin Levels

Background Serum vancomycin concentration should be carefully monitored due to its narrow therapeutic index. The prediction of serum vancomycin concentrations should be comprehensively considered, not only based on pharmacokinetic factors, such as dose, interval, and clearance but also according to patient-related factors and systemic conditions. Until now, some simulation programs and dosing calculators for therapeutic drug monitoring (TDM) have been introduced. However, due to the complexity of the pharmacokinetics and pharmacodynamics of vancomycin, they raise concerns about accuracy issues. In this study, we have developed an algorithm to predict the trough and peak concentration of vancomycin using a deep neural network (DNN). Methods Data from hospitalized patients with serum vancomycin levels between 2018 and 2022 were retrospectively reviewed, and the remaining patients’ data were analyzed after excluding subjects according to our criteria. Patients’ data were input to the TDM program and predicted vancomycin concentrations were calculated using TDM software: Pharmacokinetic Systems (PKS), version 1.10 (Copyright 1991, 1992 Abbott Laboratories, Chicago, IL, USA) and MwPharm++ (MwP; Mediware, Inc., Groningen, Netherlands). Our DNN system learned this data in the training set and based on this, the DNN system calculated the predicted values itself in the test set. The accuracy and correlation of these values compared with the actual follow-up values were analyzed statistically. Results A total of 503 patients were collected, but only 213 patients were enrolled finally. Of those, 171 patients’ data were used for the training set and 42 patients’ data were used for the testing set. As determined using Spearman’s correlation coefficients, the MwPharm++ (0.77, 95% confidence interval (CI) 0.70-0.82) had a slightly better correlation than PKS (0.81, 95% CI 0.75-0.85). And, the mean absolute error values for trough and peak vancomycin concentrations were as follows: 0.031 and 0.043 for our DNN system; 0.038 and 0.049 for the PKS system; 0.039 and 0.046 for the MwPharm++ system. Conclusion This study aimed to provide basic data to more effectively control vancomycin concentrations, increase the therapeutic effect, and minimize toxicity and side effects in clinical practice. Therefore, the use of vancomycin concentration prediction algorithms using our DNN system can overcome the limitations of current TDM programs. Further prospective studies are required to improve the accuracy of our DNN system and validate the use of DNN in clinical practice.