DeepTDM: Deep Learning-Based Prediction of Sequential Therapeutic Drug Monitoring Levels of Vancomycin

Abstract

Objective: Therapeutic drug monitoring (TDM) is essential for managing medication dosages in critically ill patients, particularly for antibiotics such as vancomycin. The dynamic physiological conditions of critically ill patients require frequent monitoring of vancomycin levels to ensure therapeutic therapeutic efficacy while minimizing toxicity. Traditional Bayesian methods and pharmacokinetic (PK) models often fail because of the complex and unpredictable nature of these patients’ conditions, as well as the limitations of standard PK modeling.Methods and procedures: This study aimed to establish a gated recurrent unit (GRU)-integrated joint multilayer perceptron network (GointMLP) model to predict sequential vancomycin TDM levels in patients in the intensive care unit. The proposed model consists of three modules to maintain consistent therapeutic vancomycin concentrations while accommodating individual patient differences. By integrating regression and classification predictions, GointMLP provides a dual mechanism for clinicians to verify the reliability of predicted values for informed decision-making. Additionally, we have developed DeepTDM, a comprehensive decision support system designed for real-time vancomycin dose optimization to enhance clinical outcomes.Results: The GointMLP provides more accurate predictions compared to traditional PK models and other machine learning/deep learning approaches. This superior performance is demonstrated not only in local validation cohorts but also in the ethnically diverse MIMIC-IV dataset, validating the model’s robust generalizability.Conclusion: This work addresses the limitations of current methodologies while leveraging advancements in deep learning techniques, particularly demonstrating the effectiveness of GointMLP in enhancing patient outcomes through precise TDM. Efforts are underway to integrate DeepTDM into clinical practice, with the anticipation that it will not only support clinicians in decision-making but also substantially improve therapeutic outcomes for patients undergoing vancomycin therapy. Clinical and Translational Impact Statement: The proposed model and software enable individualized vancomycin dosing for critically ill patients, improving precision dosing and supporting seamless integration into clinical workflows