AI-006 - DEEP LEARNING PREDICTION OF VANCOMYCIN TIME-CONCENTRATION LEVELS IN NEONATES

M. Chang¹, G. Choi², J. Koo³, H. Kim⁴, H. Lee⁴, K. Ryu⁵, J. Lim⁶, H. Kim⁶, S. Lim⁷, J. Choi¹, S. Yang⁸, S. Kang⁹, S. Lee¹; ¹Yonsei University, , , ²ImpriMedKorea, Inc, , , ³Hongik University, , , ⁴Department of Pharmaceutical Medicine and Regulatory Science, Yonsei University, , , ⁵Department of Pharmacy and Yonsei Institute of Pharmaceutical Sciences, Yonsei University, ⁶Graduate Program of Industrial Pharmaceutical Science, Yonsei University, ⁷ImpriMed, Inc.., ⁸Kyunghee University, ⁹Korea Institute of Drug Safety & Risk Management,.

Background: Vancomycin is a glycopeptide antibiotic used to treat gram-positive Staphylococci, particularly Methicillin-Resistant Staphylococcus Aureus (MRSA). Therapeutic drug monitoring (TDM) is crucial to prevent nephrotoxicity and ototoxicity and optimize its effectiveness. TDM for vancomycin is especially vital in neonates due to the critical importance of infection control for their survival. However, there is currently no study utilizing deep learning (DL) to predict vancomycin concentrations in neonates. Therefore, this research explores DL algorithms to predict vancomycin time-concentrations based on simulated data from the literature.

Methods: The study involved simulating data from population pharmacokinetics of vancomycin in neonates. The data was derived from a retrospective cohort study of vancomycin pharmacokinetics in neonates. The Mrgsolve R package (https://github.com/metrumresearchgroup/mrgsolve) was employed for data simulation. Vancomycin was administered at a dose of 5 mg/kg per dose, with dosing intervals randomly distributed between 6 and 48 hours. Body weight was simulated as a normal distribution. Vancomycin plasma concentrations were simulated for 1200 patient data points at 1-hour intervals from 0 to 96 hours. Covariates such as sex, weight, postmenstrual age (PMA), and creatinine clearance (CLcr) were considered based on the literature.

Results: Among the deep learning algorithms tested, CNN 1D performed the best in predicting vancomycin plasma concentrations between 97 and 120 hours. It achieved the lowest model loss (0.0998 for training loss, 0.1003 for validation loss, and 0.1447 for test loss) compared to other DL methods.

Conclusion: The study concludes that CNN 1D can predict vancomycin plasma concentrations for the next 24 hours after training on 96 hours of vancomycin data in neonates. This suggests the potential use of deep learning algorithms for Therapeutic Drug Monitoring (TDM) of vancomycin in this population