A Deep Learning-Based Approach for Prediction of Vancomycin Treatment Monitoring: Retrospective Study Among Patients...

Area of Science:

• Pharmacology and pharmacokinetics
• Artificial intelligence in medicine
• Critical care medicine

Background:

• Vancomycin pharmacokinetics exhibit significant variability in critically ill patients.
• Traditional population pharmacokinetic (PPK) models are population-dependent and may not meet individual patient needs.
• A deep learning system was developed to predict vancomycin therapeutic drug monitoring (TDM) levels for intensive care unit (ICU) patients.

Purpose of the Study:

• To introduce JointMLP, a novel deep-learning model for predicting vancomycin TDM levels.
• To compare the predictive performance of JointMLP against established PPK models, XGBoost, and TabNet.

Main Methods:

• A dataset of 977 cases was used for training and testing (9:1 ratio).
• External validation was performed using 1429 cases from Kangwon National University Hospital and 2394 from MIMIC-IV.
• 10-fold cross-validation and evaluation of generalization ability on MIMIC-IV data were conducted.

Main Results:

• JointMLP demonstrated superior predictive performance over PPK, XGBoost, and TabNet on both internal and external datasets.
• The model improved predictive power by up to 31% (MAE 6.68 vs 5.11) compared to PPK on the internal dataset and 81% (MAE 11.87 vs 6.56) on the external dataset.
• JointMLP showed enhanced robustness to outlier data and smaller mean errors and variances compared to PPK.

Conclusions:

• The JointMLP approach offers a valuable tool for optimizing vancomycin treatment in ICU patients.
• Improved vancomycin dosing can reduce adverse effects, bacterial resistance, and healthcare costs.
• The model's superior performance highlights its potential to assist real-world clinical decision-making.