Deep Learning Based Pharmacokinetic Model for Vancomycin

Abstract: (30 lines) Vancomycin is one of the most commonly used antimicrobial drugs in inpatient settings. National guidelines recommend Bayesian models to monitor the therapeutic drug concentration of vancomycin, especially for methicillin-resistant Staphylococcus aureus (MRSA), to minimize drug toxicity while maintaining its efficacy. Existing Bayesian models, despite being claimed as patient-specific pharmacokinetic (PK) models, use simple patient features and are studied in limited patient populations for the population-based PK parameters (the Bayesian prior). Increasingly available real-world electronic health records (EHR) provide a wide range of patient-specific data, including data on vancomycin dosage and serum levels. However, the limited flexibility of the Bayesian model structure prohibits the full use of these rich data. Deep-learning models, such as recurrent neural network (RNN), are particularly attractive for PK of vancomycin in EHR, compared to Bayesian models and other traditional machine learning models, because deep-learning models enable more flexible patient- specific inputs and possess a higher latent capacity. Thus, they deliver better predictions for a diverse population. Our deep-learning model for vancomycin (PK-RNN-V) outperforms publicly available Bayesian models but can be improved on various aspects. In Aim 1, we will improve PK-RNN-V model architectures and add more patient-specific data and a finer timestep. We will construct two-compartment PK-RNN models to increase predictive power in patients with unsteady states. We will augment PK-RNN-V with Med-BERT to improve the embedding of categorical data. We will also develop multi-track ordinary differential equations models for simultaneous prediction of serum creatinine and vancomycin levels. In Aim 2, we will use EHR from different sources to validate our PK-RNN-V model and improve the data-extraction flow and pre-processing to harmonize data from healthcare systems. We will use EHR from Houston Methodist Hospital and Memorial Hermann Hospital System/The University of Texas Health Science Center in Houston, TX, the University of Arizona in Phoenix, AZ, and the publicly available MIMIC-IV database (Boston, MA). These databases contain data from more than 121,007 patients who received at least one dose of intravenous vancomycin. In Aim 3, we will add dosing recommendations based on PK-RNN-V model predictions as a feature and validate our model in specific subgroups with challenging vancomycin PK to predict PK levels. This project will deliver substantial model improvements, leading directly to the optimization of vancomycin use in hospitals, increased in patient safety by minimizing adverse events, and reduced healthcare costs, which align with NIH’s research mission.