Abstract: TH-PO0824
Nephrocast-V2: A Pharmacokinetics-Informed Recurrent Deep Learning Model for Vancomycin Therapeutic Drug Monitoring in Critically Ill Patients
Session Information
- Pharmacology (PharmacoKinetics, -Dynamics, -Genomics)
October 22, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pharmacology (PharmacoKinetics, -Dynamics, -Genomics)
- 1900 Pharmacology (PharmacoKinetics, -Dynamics, -Genomics)
Authors
- Yousif, Zaid, University of California San Diego, La Jolla, California, United States
- Stevens, Craig A., University of California San Diego, La Jolla, California, United States
- Aronoff Spencer, Eliah, University of California San Diego, La Jolla, California, United States
- Malhotra, Atul, University of California San Diego, La Jolla, California, United States
- Nemati, Shamim, University of California San Diego, La Jolla, California, United States
Background
Vancomycin, a glycopeptide antibiotic used to treat methicillin-resistant Staphylococcus aureus infections, is challenging to dose in critically ill patients due to fluctuating renal function. We hypothesized that a recurrent deep learning model could accurately predict vancomycin concentrations up to 2 days in advance.
Methods
We trained and tested the model using electronic health record data from adult patients admitted to intensive care units (ICUs) within the University of California San Diego Health System between January 1, 2016, and June 30, 2024. Patients were included if they spent at least 24 hours in the ICU and had at least one measured vancomycin concentration. The model leveraged a Gated Recurrent Unit architecture and incorporated features including vital signs, laboratory measurements, vancomycin population elimination rate constant (k) and volume of distribution (Vd), and next-day serum creatinine (SCr) predicted by a previously developed machine learning model. The model predicted patient-specific k and Vd, which were subsequently used to calculate vancomycin concentrations over the next 2 days. Predicted concentrations were compared with measured concentrations, and predictive performance was evaluated using mean absolute error (MAE), median absolute error (MeAE), and root mean square error (RMSE) metrics.
Results
A total of 1,636 encounters met the eligibility criteria. The median age was 57.7 years, and the median ICU length of stay was 6.3 days. The median number of measured vancomycin concentrations per encounter was 2. The model achieved an MAE of 4.01 mg/L, MeAE of 2.76 mg/L, and an RMSE of 5.2 mg/L in the test set. The most important features were the predicted next-day SCr, ICU length of stay, and gender (Figure 1).
Conclusion
This study demonstrates the potential to leverage deep learning to inform vancomycin therapeutic drug monitoring. Future work will focus on the external validation of the model and on applying this framework to predict vancomycin concentrations in different patient subpopulations.
Acknowledgment
SN acknowledges grant funding from the National Library of Medicine (award number R01LM013998) and the National Institute of General Medical Sciences (award number R35GM143121).
Top 15 features ranked by importance, categorized into continuous and categorical feature groups.
Funding
- Other NIH Support