JMIR Publications

ABSTRACT

Background:

The pharmacokinetic profile of vancomycin is highly variable, especially in critically ill patients. Various deep learning models have been successful so far in the decision-making system that predicts the vancomycin therapeutic drug monitoring (TDM) level in patients in the intensive care unit (ICU).

Objective:

We aimed to establish an ideal model by comparing and integrating methods in the decision-making system that predicts the vancomycin TDM level in ICU patients.

Methods:

We proposed novel deep learning model, the joint multilayer perceptron (JointMLP) for predicting vancomycin TDM level and compared the performance of a population pharmacokinetic (PPK) model, extreme gradient boosting, and TabNet, respectively, with data collected from Dongguk University Ilsan Hospital (DUIH) and Kangwon National University Hospital (KNUH).

Results:

Of the 977 DUIH datasets, 97 were used as the test set while the remaining 879 were used as the training data set. The external validation subject was data from 1,429 KNUH subjects. For the external data set, all models were found to have significantly higher predictive power than PPK. However, in internal datasets, the JointMLP model showed significantly better performance in predicting vancomycin TDM than the PPK model (root mean squared error 8.27 vs. 10.38, P < 0.01). The JointMLP model showed better predictive performance than other models, including the PPK model, in the external dataset. The most influential variables in TDM prediction were the vancomycin volume of distribution, sex, and height in both internal and external datasets. These variables consistently had high SHAP values in both datasets

Conclusions:

JointMLP implementation for clinical use will not only provide optimal vancomycin doses but also provide many improvements in itself as the data are continuously updated, which will result in continued increments in accuracy.