Abstract
INTRODUCTION: Vancomycin is a critical antibiotic for treating methicillin-resistant Staphylococcus aureus and other gram-positive bacterial infections, but achieving and maintaining therapeutic trough concentrations is challenging. OBJECTIVES: We hypothesized that a deep learning model could accurately predict vancomycin trough concentrations 2 days in advance in critically ill patients and provide recommendations for optimal dosing adjustments to achieve target drug concentrations. METHODS: We trained and validated the model using electronic health record (EHR) data from adults admitted to the University of California San Diego Health system intensive care units (ICUs) from January 1, 2016, to June 30, 2024. Features included patient demographics, comorbidities, vital signs, laboratory measurements, medications, and vancomycin dosing information. The model architecture combined Long Short-Term Memory and Multi-Head Attention layers, supplemented with skip connections to incorporate past dosage information at the final layer of the deep learning model. Model performance was evaluated using mean absolute error (MAE) and root mean square error (RMSE) metrics. RESULTS: A total of 2205 encounters met the eligibility criteria. The median age was 57 years, and the median ICU length of stay was 4.9 days. The model achieved an MAE of 3.15 mg/L and an RMSE of 4.17 mg/L, comparable to that of a critical care pharmacist aided by a Bayesian dosing software. Additionally, deviations from patient-specific model-based dose recommendations were generally associated with nontherapeutic vancomycin levels. CONCLUSION: This study demonstrates the potential to leverage deep learning to individualize and support vancomycin therapeutic drug monitoring in critically ill patients.