Predicting Ketamine Exposure in Pediatric Status Epilepticus: A Pharmacokinetic-Machine Learning Approach.

Predicting Ketamine Exposure in Pediatric Status Epilepticus: A Pharmacokinetic–Machine Learning Approach - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Eur J Drug Metab Pharmacokinet . 2026 Apr 3;51(3):327–334. doi: 10.1007/s13318-026-00996-1 Predicting Ketamine Exposure in Pediatric Status Epilepticus: A Pharmacokinetic–Machine Learning Approach Adeboye O Bamgboye 1, 2, ✉ , Lisa D Coles 1, 2 , Sílvia M Illamola 1 , Julian Wolfson 3 , James C Cloyd 1, 2 Author information Article notes Copyright and License information 1 Department of Experimental and Clinical Pharmacology, College of Pharmacy, McGuire Research Translational Facility, University of Minnesota, Minneapolis, MN 55455 USA 2 Center for Orphan Drug Research, College of Pharmacy, University of Minnesota, Minneapolis, MN USA 3 Division of Biostatistics and Health Data Science, School of Public Health, University of Minnesota, Minneapolis, MN USA ✉ Corresponding author. Received 2026 Jan 12; Accepted 2026 Mar 17; Issue date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which permits any non-commercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc/4.0/ . PMC Copyright notice PMCID: PMC13221361  PMID: 41933112 Abstract Background and Objective Status epilepticus is a life-threatening neurological emergency. Ketamine combined with levetiracetam is a promising therapy being investigated for established status epilepticus. Early drug exposure estimates could inform dosing, but intensive pharmacokinetic sampling is challenging in pediatric emergency settings due to logistical and ethical constraints. Existing population pharmacokinetic methods with Bayesian forecasting require multiple samples per individual, which is difficult in this context. The objective of this work was to develop a framework that utilizes population pharmacokinetic-simulated datasets as inputs for machine learning models to predict early ketamine exposure in pediatric status epilepticus. Methods We simulated 100,000 concentration-time profiles with covariates using a published population pharmacokinetic model. Additional predictor variables were simulated using distributions from a previously published status epilepticus trial. The simulated data were used to train machine learning algorithms to predict early exposure (AUC 0-2h ) using two samples per individual. The machine learning models were validated internally. Results We evaluated four machine learning algorithms, including LASSO, random forest (RF), K-nearest neighbor (KNN), and gradient boosted machine (GBM). The ensemble models (GBM and RF) achieved superior performance, with internal validation results indicating no overfitting. The root mean square errors for RF and GBM were 0.183 mg h/L and 0.185 mg h/L respectively, while mean absolute errors of RF and GBM were 0.103 mg h/L and 0.102 mg h/L, respectively Conclusions We demonstrate the feasibility of developing population pharmacokinetic-informed machine learning models to accurately predict early ketamine exposures in pediatric status, where pharmacokinetic sampling is a challenge. Supplementary Information The online version contains supplementary material available at 10.1007/s13318-026-00996-1. Key Points Early drug exposures could inform dosing decisions for pediatric status epilepticus management. Logistical and ethical factors limit intensive pharmacokinetic sampling in pediatric emergent conditions. A hybrid machine learning-population pharmacokinetic approach is a feasible method of predicting early ketamine exposures from sparse sampling in pediatric status epilepticus. Open in a new tab Introduction Pharmacokinetic (PK) studies in children with emergent conditions such as status epilepticus (SE) are challenging for several reasons, including practical limitations regarding the volume of blood and number of samples that can be safely collected, concerns about unnecessary phlebotomy, and cost [ 1 , 2 ]. Many standard pharmacokinetic studies rely on intensive sampling [ 3 ], which involves the collection of several blood samples to estimate a drug’s PK parameters. For many pediatric studies, intensive sampling is often not possible, thus necessitating innovative approaches that can maximize the information obtained from sparse data. Population pharmacokinetic (PopPK) modeling is an established non-linear mixed effects statistical method capable of utilizing information from sparse sampling at opportunistic schedules, which are typical in emergent conditions like SE [ 4 – 6 ]. PopPK utilizes statistical methods to estimate population pharmacokinetic parameters such as clearance even with sparse data. Furthermore, PopPK can characterize the within-subject, between-subject variability, and residual error [ 6 ]. This process is however resource-intensive requiring a lot of time and expertise [ 7 ]. In addition, ML models have better flexibility compared to standard PopPK models when more data becomes available since the latter are usually constrained by the model structure. Clinical response for emergent conditions like SE often relies on early drug exposures which can be challenging to estimate with sparse sampling. Extensive early sampling is frequently precluded by the care needed to stabilize and monitor patients. Abhishek et al. showed that PopPK can provide reliable estimates of early drug exposure even with limited sampling, but this approach retains the limitations of standard PopPK methods [ 5 ]. Machine learning (ML) methods are becoming increasingly popular for drug development [ 7 – 9 ] especially with the growing complexities of drug discovery and development and a deeper understanding of sources of variability in drug response. ML methods typically involve training algorithms to learn relationships within a dataset and maximize the ability of the algorithm to classify, predict, cluster, or identify patterns in a new dataset [ 10 , 11 ]. Several studies have demonstrated the benefit of combining ML methods with traditional PopPK methods to predict pharmacokinetic parameters, and drug concentrations [ 12 – 14 ] with results superior to those of either method alone. These studies employ ML algorithms such as artificial neural networks (ANNs) to predict plasma concentrations, distribution kinetics, and clearance of antibiotics and anesthetics, primarily in adult populations. To the best of our knowledge, a hybrid PopPK-ML approach has not been used for early drug exposure prediction in children with emergent conditions. The aim of this study was to investigate the feasibility of developing a hybrid PopPK-ML framework for predicting early ketamine exposure (AUC 0-2h ) in pediatric SE patients from two early concentrations. Methods Published Models used for Simulation After carefully reviewing published ketamine PopPK models in children, the PopPK model by Sherwin et al. [ 15 ] was selected and used to simulate the individual ketamine concentration-time data for n = 100,000 children. This model by Sherwin et al. is a 2-compartment PK model developed from concentration–time data in children who received a 1–1.5 mg/kg intravenous bolus of ketamine. The number of individual profiles to simulate was chosen to provide larger training datasets, improve statistical precision, and incorporate wider interindividual variability. The interindividual variability ranged from 34.1% for the peripheral volume to 97.6% for the central volume. The proportional residual unexplained variability (RUV) was fixed at 20%, while the additive RUV was fixed to zero. We simulated demographic and seizure duration data using the Established Status Epilepticus Treatment Trial (ESETT) report [ 16 ]. Demographic information included age (years), weight (kg), race, sex, and seizure duration. Age was restricted between 1 and 18 years. Age and weight were simulated from a truncated normal distribution using the truncnorm package in R, and assigned to three age groups (Infants and toddlers: 1–5 years; young children: 6–10 years; and adolescents: 11–18 years). In the simulated dataset, the sex distribution was 53% male and 47% female. We used a uniform dose of 2 mg/kg, which is consistent with doses typically administered to children receiving ketamine off-label for status epilepticus treatment [ 17 ]. Including covariates served to increase interindividual variability [ 18 ]. The PopPK parameters of the reference model are detailed in Table S1. Simulation and Modeling Based on ketamine’s pharmacokinetics, the drug exposure metric chosen was the area under the curve in the first 2 h (AUC 0-2h ) to reflect early drug exposures. The schematic of the simulation methodology is shown in Fig. 1 . The steps are discussed in detail below Fig. 1. Open in a new tab Schema of the methodology used. AUC area under the curve, ESETT Established Status Epilepticus Treatment Trial, GBM gradient boosted trees, KNN K-nearest neighbor, LASSO least absolute shrinkage and selection operator, ML machine learning, PopPK population pharmacokinetic, RForest random forest. Figure Created in BioRender. Bamgboye, A. (2025) https://BioRender.com/g66ag0m Step 1: Simulation of concentration-time profiles without RUV Using the PopPK model by Sherwin et al. [ 15 ], rich concentration-time profiles without RUV were simulated for 100,000 children receiving 2 mg/kg of ketamine. Simulated profiles were generated using R (v.4.3.1) and the mrgsolve package. Step 2: Demographic and seizure duration data Patient demographic and seizure duration data were simulated for 100,000 individuals using distributions from the ESETT study [ 16 ] in R (v.4.3.1) with the truncnorm package. Data from steps 1 and 2 were merged to form the reference dataset. The creation of additional predictors to add information was performed as previously described [ 18 ]. Step 3: Reference AUC calculation The PopPK-simulated profiles generated in step 1 were used to calculate the individual reference AUC 0-2h by integrating the concentrations in the central compartment from the start of drug dosing to time t = 2 h. These reference AUCs were withheld and used as ground truth for calculating the prediction accuracy (internal validation) of our ML models. Step 4: Simulation of concentration-time profiles with RUV Using the PopPK model by Sherwin et al. [ 15 ], rich concentration–time profiles (0–2 h) with RUV fixed at 20% were simulated for 100,000 children receiving 2 mg/kg of ketamine. The simulated concentration–time profiles were merged with patient demographics to create a prediction dataset containing concentration and covariates. Step 5: ML AUC calculation Data from the prediction dataset were used to train four classical ML models: least absolute shrinkage and selection operator (LASSO), gradient boosted machine (GBM), random forest (RForest), and K-nearest neighbor (KNN). Two concentrations (CP1 and CP2) at time = t1 (10 min post-dose) and time = t2 (60 min post-dose), respectively, alongside age, race, sex, weight, and seizure duration, were used as predictors in each ML model. The two time points (t1 and t2), were chosen to accommodate clinically feasible sampling windows for pediatric SE while ensuring sufficient data to capture both distribution and elimination phases of ketamine. To account for sampling time variability, we incorporated buffers (± 5 min around t1; ± 15 min around t2). Candidate predictors were selected based on biological plausibility and their influence on ketamine pharmacology. The outcome variable was the AUC 0-2h . The predicted AUCs were called ‘ML AUCs.’ The ML training, predictions, and evaluation were performed using R (v.4.3.1). Step 6: Predictive performance evaluation The root mean square error (RMSE) and mean absolute error (MAE) were calculated for the four ML methods : R M S E = ∑ y i - y ^ i 2 n M A E = ∑ y i - y ^ i n where y i refers to observed values and ŷ i refers to the model-predicted values. Finally, the coefficient of determination ( R 2 ) values were calculated for each ML model. Development of ML Models Tuning parameters included the tunelength for KNN (tunelength, starting value set to 10), lambda for LASSO (automatically tuned), number of trees, interaction depth, and distribution for GBM, and the number of variables sampled at each split was set to the default for RForest. Model training was performed with 5-fold or 10-fold cross-validation to tune the hyperparameters and prevent overfitting. The parameter combinations with the lowest RMSE were selected for the KNN, GBM, and RForest models. The optimal lambda was selected based on the 1-standard error rule for LASSO. Results Simulated Patient Data The characteristics of the training and test datasets are reported in Table 1 . The training and test datasets included 100,000 simulated individuals each. The distribution of sex and race was similar between training and test datasets, as were the medians of the plasma concentrations at the two sampling times, i.e., CP1 and CP2. The outcome also exhibited similar median values and variability in both datasets. Taken together, the similar distributions support reliable model-performance comparison. Table 1. Characteristics of the features in the training and test datasets Train ( n = 100,000) Test ( n = 100,000) Age, years 5 [1–18] 5 [1–18] Sex ( n , %) Male 52,992 (53) 52,992 (53) Female 47,008 (47) 47,008 (47) Weight, kg 17.3 [10.0–65.0] 17.3 [10.0–65.0] Seizure duration, min 62.9 [0.0–204.8] 62.9 [0.0–204.8] Race Black 36,826 (36.8) 36,826 (36.8) Hispanic 13,856 (13.9) 13,856 (13.9) White 36,998 (37.0) 36,998 (37.0) Other 12,320 (12.3) 12,320 (12.3) First plasma concentration (CP1), mg/L 1.04 [0.002–86.73] 1.05 [0.003–96.91] Second Plasma concentration (CP2), mg/L 0.04 [0.000–0.72] 0.04 [0.000–0.93] AUC 0-2 , h mg/L 0.39 [0.05–8.39] 0.39 [0.04–8.92] Open in a new tab Data are presented as median [interquartile range] for continuous data and n (%) for categorical data AUC 0-2 area under the curve in the first 2 h following ketamine administration Feature Importance We assessed variable importance using either a total reduction in mean squared error or permutation-based methods. The analysis revealed consistent patterns across RForest, GBM, and KNN. Weight and the first plasma sample (representing the earliest ketamine concentration after drug administration) were the most important predictors of AUC 0-2h . Age had a modest contribution in the RForest model, but not the other two models (GBM and KNN). We also examined the magnitudes of the LASSO coefficients. The two plasma concentrations, and weight were retained as essential predictors matching the overall trend seen in other ML models. Model Performance The evaluation metrics results for each ML method are summarized in Table 2 . The RMSE is used to evaluate the magnitude of error between predicted and observed values. RMSE values were similar for the ensemble methods. The RForest model had the smallest RMSE (0.183 mg h/L) followed by GBM (0.185 mg h/L), LASSO (0.216 mg h/L), and KNN (0.221 mg h/L). The MAE also measures average magnitude of errors, but it treats all errors equally as opposed to RMSE which penalizes large errors. The best performing ML model based on MAE values was GBM (0.102 mg h/L), followed closely by RForest (0.103 mg h/L), LASSO (0.125 mg h/L), and KNN (0.130 mg h/L). The LASSO, GBM, and RForest models could explain more than 70% of the variability seen in the data ( R 2 = 0.71, 0.78, and 0.78, respectively). Table 2. Model performance for prediction of AUC 0-2h Prediction models Evaluation metric LASSO KNN GBM RForest RMSE 0.216 (0.215–0.217) 0.221 (0.220–0.222) 0.185 (0.184–0.186) 0.183 (0.182–0.183) MAE 0.125 (0.124–0.126) 0.130 (0.128–0.131) 0.102 (0.101–0.103) 0.103 (0.102–0.104) R 2 0.71 0.68 0.78 0.78 Open in a new tab GBM gradient boosted machine, KNN K-nearest neighbor, LASSO least absolute shrinkage and selection operator, MAE mean absolute error,, R 2 coefficient of determination, RForest random forest, RMSE root mean square error Internal validation analysis showed minimal differences in model performance for the training and test datasets suggesting that there was no overfitting. The model performance for the training dataset is shown in the Supplementary Table S2. The calibration box plots showing calibration performance across predicted AUC 0-2h quartiles are shown in Fig. 2 . Fig. 2. Open in a new tab Variable importance plots for machine learning models predicting AUC 0-2h . A gradient boosting machine, B random forest, C K-nearest neighbor. Variable importance scores were calculated based on total reduction in mean squared error A or permutation-based methods ( B and C ). Variables are ranked from most to least important For all four ML models, the mean observed value for each quartile align closely with the median predicted values particularly at lower exposure levels indicating acceptable calibration. Discussion In this study, we show that ML models trained on two PopPK-simulated concentrations per child accurately predicted early ketamine exposure demonstrating the feasibility of this hybrid framework. Previous studies have demonstrated the benefits of hybrid PopPK-ML methods [ 19 ]. For example, Keutzer et al. [ 20 ] showed that hybrid PopPK–ML approaches can accurately predict plasma concentrations and drug exposures. However, studies demonstrating the feasibility of these hybrid frameworks in pediatric populations with emergent neurological conditions are lacking. This is the first study showing the utility of ML-aided early exposure predictions in children with an emergent condition. The ensemble methods (i.e., RForest and GBM) performed better than the other two ML (KNN and LASSO) methods as evidenced by the lower RMSE and MAE values and higher R 2 values. Ensemble methods are highly powerful algorithms that rely on combining several individual methods to reduce prediction errors [ 21 ] and have been widely used in clinical prediction tasks [ 22 ]. The improved prediction accuracy demonstrated in our study further supports their ability to capture complex biological relationships. While age was an important predictor of AUC 0-2h (Fig. 3 ), age and weight are physiologically correlated. We decided to retain both variables because the primary study objective was to demonstrate the feasibility of the approach rather than identify independent predictors of early drug exposure. Fig. 3. Open in a new tab Calibration boxplots across predicted AUC 0-2h quartiles for A LASSO, B K-nearest neighbor, C gradient boosting machine, and D random forest. Boxplots show the distribution of predicted values within each quartile and the red dots represent the mean observed AUC 0-2h for each quartile Early drug exposure is plausibly linked to clinical response and safety outcomes, especially for emergent conditions like SE. However intensive PK sampling is rarely feasible in the emergency department. This creates a dilemma in which early drug exposure estimates could be used to guide dosing decisions, but sparse sampling dictated by the clinical reality provides insufficient data to reliably estimate individual PK parameters. The hybrid PopPK–ML framework addresses this gap by integrating the information obtained from PopPK-simulated datasets to train ML algorithms. The result is a tool that learns to predict early exposure from the limited pharmacokinetic data likely to be obtained in practice. Early therapeutic exposures are typically important for producing a drug response in emergent conditions such as status epilepticus in which attaining rapid target concentrations is important [ 5 ] . The number of samples required depends on the PK properties of the drug. Ketamine has a rapid distribution and elimination informing our decision to select two time points to maximize information about the distribution and elimination phases [ 23 ]. For drugs with slower elimination, we could utilize one drug concentration to estimate early dug exposure which is mostly informed by distribution [ 5 ]. This study is not intended to be a comparison of PopPK and ML predictions. Previous studies have shown that ML predictions may be superior to those of PopPK under certain conditions [ 12 ]. In our study, constraints were imposed on the sampling times. Often, PopPK models are developed with concentrations obtained from more frequent sampling to provide reliable PK parameter estimates. Similarly, the concentrations used as predictors for the ML models were derived from a published PopPK model limiting the direct comparison of ML and PopPK methods. Further studies are needed to directly compare ML and Bayesian PopPK methods in the population under study. PopPK can capture and partition different sources of variability: inter-individual variability, intra-occasion, and residual error using a nonlinear mixed effects framework with sparse data [ 6 ]. The final PopPK model can then be used to simulate sufficient training datasets. In contrast, ML approaches often do not partition variance in the same way, and as such may struggle with sparse training data [ 20 ]. A hybrid PopPK–ML framework combines the mechanistic framework of PopPK with the flexibility of ML to improve predictions in pediatric emergency settings. There are several limitations of our study. The main limitation is the lack of clinical data for external validation of the model. External validation is limited by the paucity of pediatric SE studies and challenges in pharmacokinetic sampling in this population. This work serves as a proof-of-concept for a PopPK-informed ML model. Further studies, particularly multicenter studies would enable validation of this model. The assumption that the simulated population accurately reflects the literature populations in ESETT and the literature PopPK model by Sherwin et al.'s study has its shortcomings. Nevertheless, simulations from published PopPK models are increasingly being accepted in pediatric drug development pharmacokinetic studies [ 24 ]. We did not incorporate optimal sampling design which could further optimize timepoint selection for sparse sampling to maximize accuracy. Instead, we evaluated clinically feasible sampling windows and incorporated buffers (± 5–15 mins) to account for deviations in sampling time inherent in emergency care. Furthermore, while the simulated sample size of 100,000 enables robust model training, such data collection is unlikely to be feasible in real-world settings for ethical and logistic reasons. Nevertheless, incorporating inter-individual variability and residual error in the simulations helps account for the heterogeneity expected in real pediatric populations. We did not exhaust the library of machine learning methods. Some powerful ML methods like ANNs, extreme gradient boosting (XGBoost), and support vector machine have been shown to perform well for pharmacokinetic predictions in other studies [ 12 ]. In addition, the simulated demographic characteristics were based on the ESETT trial population. Clinical covariates such as concomitant medications, or organ function were not included in this analysis, but could be part of future implementations. The inclusion of features such as seizure duration and sex, that are not mechanistically linked to ketamine exposures served as a means to demonstrate how to incorporate readily available clinical data into the hybrid PopPK–ML framework for future application. Conclusion We have shown that ML methods are useful companion tools to standard PopPK methods for predicting early drug exposures in children with an emergent condition. Ensemble methods were superior to other ML methods (KNN and LASSO) in our study and were able to predict early ketamine exposure with acceptable precision. Hybrid PopPK–ML methods could help enhance pharmacokinetic studies focusing on the relationships between early drug exposure and treatment outcomes in emergent conditions like status epilepticus. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 18 KB) (18.1KB, docx) Supplementary file2 (RMD 27 KB) (27.2KB, rmd) Author Contributions AOB, LDC, SMI, JW, and JCC wrote the manuscript. AOB, and JW designed the research. AOB. performed the research. AOB, LDC, JW, and SMI analyzed the data. Funding No funding was received for this work Data Availability The datasets generated and used in this study are synthetic data created through pharmacokinetic simulations. Code Availability The complete code for data generation, model development, and analysis is provided in the supplementary materials. Declarations Conflict of interest AOB, LDC, SMI, JW, and JCC report no conflicts of interests. Ethical Approval Not applicable. Consent to Participate Not applicable. Consent for Publication Not applicable. References 1. Barker CIS, Standing JF, Kelly LE, Hanly Faught L, Needham AC, Rieder MJ, et al. Pharmacokinetic studies in children: recommendations for practice and research. Arch Dis Child. 2018;103:695–702. 10.1136/archdischild-2017-314506. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Orito Y, Kakara M, Okada A, Nagai N. Model-based approach to sampling optimization in studies of antibacterial drugs for infants and young children. Clin Transl Sci. 2021;14:1543–53. 10.1111/cts.13018. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Ariano RE, Duke PC, Sitar DS. The influence of sparse data sampling on population pharmacokinetics: a post hoc analysis of a pharmacokinetic study of morphine in healthy volunteers. Clin Ther. 2012;34:668–76. 10.1016/j.clinthera.2012.01.023. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Autmizguine J, Benjamin DK, Smith PB, Sampson M, Ovetchkine P, Cohen-Wolkowiez M, et al. Pharmacokinetic studies in infants using minimal-risk study designs. Curr Clin Pharmacol. 2014;9:350–8. 10.2174/1574884709666140520153308. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Sathe AG, Brundage RC, Ivaturi V, Cloyd JC, Chamberlain JM, Elm JJ, et al. A pharmacokinetic simulation study to assess the performance of a sparse blood sampling approach to quantify early drug exposure. Clin Transl Sci. 2021;14:1444–51. 10.1111/cts.13004. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Mould DR, Upton RN. Basic concepts in population modeling, simulation, and model-based drug development—Part 2: introduction to pharmacokinetic modeling methods. CPT Pharmacomet Syst Pharmacol. 2013;2:e38. 10.1038/psp.2013.14. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Zhu X, Huang W, Lu H, Wang Z, Ni X, Hu J, et al. A machine learning approach to personalized dose adjustment of lamotrigine using noninvasive clinical parameters. Sci Rep. 2021;11:5568. 10.1038/s41598-021-85157-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Richardson S, Irurzun Arana I, Nowojewski A, Zhou D, Leander J, Tang W, et al. A machine learning approach to population pharmacokinetic modelling automation. Commun Med. 2025;5:327. 10.1038/s43856-025-01054-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Meng H-Y, Jin W-L, Yan C-K, Yang H. The application of machine learning techniques in clinical drug therapy. Curr Comput Aided Drug Des. 2019;15:111–9. 10.2174/1573409914666180525124608. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Sarker IH. Machine learning: algorithms, real-world applications and research directions. SN Comput Sci. 2021;2:160. 10.1007/s42979-021-00592-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Reichstein M, Camps-Valls G, Stevens B, Jung M, Denzler J, Carvalhais N, et al. Deep learning and process understanding for data-driven earth system science. Nature. 2019;566:195–204. 10.1038/s41586-019-0912-1. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Ogami C, Tsuji Y, Seki H, Kawano H, To H, Matsumoto Y, et al. An artificial neural network−pharmacokinetic model and its interpretation using Shapley additive explanations. CPT Pharmacomet Syst Pharmacol. 2021;10:760–8. 10.1002/psp4.12643. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Nigo M, Tran HTN, Xie Z, Feng H, Mao B, Rasmy L, et al. PK-RNN-V E: a deep learning model approach to vancomycin therapeutic drug monitoring using electronic health record data. J Biomed Inform. 2022;133:104166. 10.1016/j.jbi.2022.104166. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Ingrande J, Gabriel RA, McAuley J, Krasinska K, Chien A, Lemmens HJM. The performance of an artificial neural network model in predicting the early distribution kinetics of propofol in morbidly obese and lean subjects. Anesth Analg. 2020;131:1500. 10.1213/ANE.0000000000004897. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Sherwin CMT, Stockmann C, Grimsrud K, Herd DW, Anderson BJ, Spigarelli MG. Development of an optimal sampling schedule for children receiving ketamine for short-term procedural sedation and analgesia. Paediatr Anaesth. 2015;25:211–6. 10.1111/pan.12521. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Chamberlain JM, Kapur J, Shinnar S, Elm J, Holsti M, Babcock L, et al. Efficacy of levetiracetam, fosphenytoin, and valproate for established status epilepticus by age group (ESETT): a double-blind, responsive-adaptive, randomised controlled trial. Lancet. 2020;395:1217–24. 10.1016/S0140-6736(20)30611-5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Adhikari A, Yadav SK, Nepal G, Aryal R, Baral P, Neupane P, et al. Use of ketamine in super refractory status epilepticus: a systematic review. Neurol Res Pract. 2024;6:33. 10.1186/s42466-024-00322-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Woillard J-B, Labriffe M, Prémaud A, Marquet P. Estimation of drug exposure by machine learning based on simulations from published pharmacokinetic models: the example of tacrolimus. Pharmacol Res. 2021;167:105578. 10.1016/j.phrs.2021.105578. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Li G, Sun Y, Zhu L. Application of machine learning combined with population pharmacokinetics to improve individual prediction of vancomycin clearance in simulated adult patients. Front Pharmacol. 2024. 10.3389/fphar.2024.1352113. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Keutzer L, You H, Farnoud A, Nyberg J, Wicha SG, Maher-Edwards G, et al. Machine learning and pharmacometrics for prediction of pharmacokinetic data: differences, similarities and challenges illustrated with rifampicin. Pharmaceutics. 2022;14:1530. 10.3390/pharmaceutics14081530. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Naderalvojoud B, Hernandez-Boussard T. Improving machine learning with ensemble learning on observational healthcare data. AMIA Annu Symp Proc. 2024;2023:521–9. [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Mahajan P, Uddin S, Hajati F, Moni MA. Ensemble learning for disease prediction: a review. Healthcare (Basel). 2023;11:1808. 10.3390/healthcare11121808. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Peltoniemi MA, Hagelberg NM, Olkkola KT, Saari TI. Ketamine: a review of clinical pharmacokinetics and pharmacodynamics in anesthesia and pain therapy. Clin Pharmacokinet. 2016;55:1059–77. 10.1007/s40262-016-0383-6. [ DOI ] [ PubMed ] [ Google Scholar ] 24. Dorlo TPC, Svensson EM. Pharmacokinetic-pharmacodynamic modeling and simulation in clinical practice and studies. In: Macheras P, editor. Advances in Pharmacokinetics and Pharmacodynamics. Cham: Springer International Publishing; 2023. p. 109–35. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary file1 (DOCX 18 KB) (18.1KB, docx) Supplementary file2 (RMD 27 KB) (27.2KB, rmd) Data Availability Statement The datasets generated and used in this study are synthetic data created through pharmacokinetic simulations. The complete code for data generation, model development, and analysis is provided in the supplementary materials. Articles from European Journal of Drug Metabolism and Pharmacokinetics are provided here courtesy of Springer ACTIONS View on publisher site PDF (1.6 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Detecting neuroplastic effects induced by ketamine in healthy human subjects: A multimodal approach#424672
• Ketamine in Patients Undergoing TEVAR Procedures Receiving NCI#377328
• Monitoring of dual-drug combination therapy in pediatric epilepsy patients: a machine learning model for simultaneous VPA-LEV concentration-dose prediction.#165629
• Effect of Ketamine on Laboratory-induced Stress in Healthy Subjects#848117
• Proper Dosage of Ketamine in Intravenous Regional Anesthesia#1059868