Background: Procalcitonin is a widely used biomarker that can support infection diagnosis. Procalcitonin increases with infection but also following non-infection events such as birth and surgery thus complicating interpretation. A population model that describes the time course of procalcitonin concentrations has been developed in a population ranging from birth to 19 weeks of age and actioned as an online tool (https://aucklandpharmacometrics.shinyapps.io/PCT_No_Infection/). The tool provides patient-specific reference ranges of expected procalcitonin concentrations which may help to guide whether the observed concentration may indicate infection.

Aims: Undertake a limited external validation of the clinical tool.

Methods: Data from a pilot study of 16 neonates admitted to the Neonatal Intensive Care Unit in New Zealand were available. Fifty-six procalcitonin concentrations and the clinician's decision of infection status were recorded. Concentrations from each infected or non-infected period were compared with predictions from the tool.

Results: Most neonates (62.5%) were preterm (gestational age median: 28 weeks; 95%CI: 23.8 - 40.0). Thirteen neonates had suspected infection and received antibiotics from birth for at least 36 hours. Of these, seven were non-infected, five were infected, and one was initially non-infected but became infected within two weeks. The tool was unable to identify neonates who had an infection at birth (1/5 correct). However, it correctly identified 91% (10/11) of the non-infected periods.

Conclusion: Caution is required when using this tool to identify neonates known to be infected on the day of birth, but it performs well for identifying non-infected neonates in the following days.

Keywords: procalcitonin, neonates, infection

Background:
Pharmacokinetic (PK) population models can be used to estimate tacrolimus area-under-the-curve (AUC) using limited sampling strategies (LSS). Microsampling approaches now makes it feasible to obtain a full PK-profile and determine AUC. While combining these methods is appealing, the clinical applicability of using LLS developed on microsampling has never been evaluated.

Aim:
The aim of the Man versus Machine study is to compare the performance of model-based strategies for tacrolimus AUC estimation to gold standard trapezoidal AUCs using a dataset of microsample full PK profiles in solid organ transplant patients.

Method:
Five different model-based strategies aiming at estimating AUCs from a LSS (0,1,3 hours for immediate-release and extended-release tacrolimus and 0,8,10 hours for LCP-tacro) were blindly tested.
AUCs estimated using model-based strategies were then compared with the corresponding observed reference trapezoidal AUCs. Model performances were deemed appropriate if less than 10% of profiles were estimated with a bias above 20%.

Results:
One-hundred and twenty-one patients were included in the study. The proportion of individual AUCs with bias >20% (failure) for each five models were: 6.6, 6.6, 7.1, 9.9 and 15.3%, respectively, meaning that between 1/6 and 1/15 has an inappropriate AUC estimation with the LSS approach.

Conclusion:
Four of five the five tested models displayed appropriate estimation of AUC when compared with reference AUCs using microsampling. A combined approach using limited microsampling strategy with PK modeling appears clinically feasible in most cases for tacrolimus AUC monitoring.

Keywords: transplantation; immunosuppressants; pharmacokinetics; drug monitoring

Background:
Predicting vancomycin exposure is critical for optimizing dosing regimens in sepsis patients. While population pharmacokinetic (PPK) models are commonly used, their performance is limited. Machine learning (ML) models may offer advantages, but it is unclear which model—PPK, Bayesian, ML, or hybrid PPK-ML—best predicts vancomycin exposure across clinical scenarios.

Aims:
This study compares the performance of PPK, Bayesian, ML, and hybrid PPK-ML models in predicting the 24-hour area under the blood concentration curve (AUC24) to support precision dosing in sepsis.

Methods:
Data from sepsis patients treated with intravenous vancomycin were sourced from the MIMIC-IV database. The dataset was split into training and testing sets to develop the four models. AUC24 was predicted using all models, and performance was evaluated with mean absolute error, mean squared error, root mean squared error, mean absolute percentage error (MAPE), and R².

Results:
A total of 4,059 patients were included. In the absence of vancomycin concentration data, the hybrid model outperformed PPK and Bayesian models, with MAPE improvements of 58% and 17%, respectively. When concentration data were available, the Bayesian model performed best (MAPE: 13.37% vs. 68.17%, 34.17%, and 28.52% for PPK, Random Forest, and hybrid models).

Conclusions:
The hybrid model is recommended when concentration are unavailable, while the Bayesian model is preferred when concentration are available, offering robust strategies for precise vancomycin dosing in sepsis patients.

Keywords: Population pharmacokinetics, Machine learning, Sepsis, Vancomycin, Drug exposure

Background: Vancomycin-associated acute kidney injury (AKI) continues to pose a major challenge to both patients and healthcare providers.

Aims: The purpose of this study is to construct a machine learning framework for stratified predicting and interpreting vancomycin-associated AKI.

Methods: This study encompasses an analysis of the medical records of 724 patients who received vancomycin therapy and underwent therapeutic drug monitoring between January 1, 2015, and September 30, 2020. The basic clinical information, vancomycin dosage and days, comorbidities and medication, laboratory indicators of the patients were recorded. Machine learning algorithm of XGBoost was used to construct a series risk prediction model for vancomycin-associated AKI in different underlying diseases.

Results: The vast majority of sub-model performed best on the corresponding sub-dataset. Additionally, the aim of this study was to explain each model and to explore the influence of clinical variables on prediction. As the results of the analysis showed that in addition to the common indicators (serum creatinine and creatinine clearance rate), some other underappreciated indicators such as serum cystatin and cumulative days of vancomycin administration, weight and age, neutrophils and hemoglobin were the risk factors for cancer, diabetes mellitus, heptic insufficiency respectively. Stratified analysis of the comorbidities in patients with vancomycin-associated AKI further confirmed the necessity for different patient populations to be studied.

Conclusions: This study successfully developed a machine learning model capable of predicting the risk of vancomycin-associated AKI and identifying its potential risk factors.

Key words: acute kidney injury; machine learning; risk stratification; stratification analysis; vancomycin.

Background: Bedaquiline (BDQ) has shown great value in the treatment of multidrug-resistant tuberculosis (MDR-TB) in recent years. However, large pharmacokinetic variability was observed in patients following standard BDQ dosing. BDQ is also reported to induce cardiac QT interval prolongation.

Aims: This study aims to perform a population pharmacokinetic-pharmacodynamic (PPK-PD) analysis of BDQ in Chinese MDR-TB patients to enable model-based precision dosing and investigate the concentration-QTc relationships of BDQ.

Methods: Blood samples and therapeutic drug monitoring (TDM) results of BDQ were collected for PK evaluation, and ECGs were recorded for QTc calculation at the designed timepoints. The PPK-PD model was developed using nonlinear mixed effects modeling (NONMEM) software. An effect compartment model was utilized to explain the relationship between BDQ concentrations and ∆QTc.

Results: A total of 211 BDQ concentrations and 350 QTc measurements were obtained from 105 Chinese MDR-TB patients. The PK characteristics of BDQ was adequately described using a two-compartment model with first-order absorption and elimination. The estimated apparent clearance and volume of distribution were 1.77 L/h and 52.4 L, respectively. Body weight and albumin were significant covariates for clearance of BDQ, with body weight also affecting the apparent volume of distribution. ∆QTc exhibited a linear relationship with BDQ concentration. A linear model properly characterized BDQ-induced QTc prolongation.

Conclusions: A PPK-PD model based on sparse data was established for prediction of BDQ exposure in patients with MDR-TB. TDM and PPK modeling provide valid evidence on the precision dosing of bedaquiline.

Keywords: Bedaquiline; PPK-PD model; concentration-QTc relationship; multidrug-resistant tuberculosis; TDM

Background: The pharmacokinetics of linezolid can vary significantly, particularly in critically ill patients. Model-informed precision dosing (MIPD) enables dose individualization by employing one or several population pharmacokinetic (popPK) models to optimize drug exposure, thereby enhancing therapeutic success and minimizing adverse effects.

Aims: To compare the performance of MIPD of linezolid using a single-model and model averaging algorithm (MAA) approach.

Methods: The predictive performance (PP) of popPK models of linezolid (n=30) was externally evaluated. Theoretical target attainment (TTA) was selected as the indicator of PP, demonstrating the percentage of expected trough concentrations within 2-8 mg/L when using popPK models for MIPD. MAA was analyzed in TDMx for two groups. Group one comprised models exhibiting the highest and lowest TTA (three each). In group two, the three models with the highest tendency to overpredict or underpredict were included, respectively.

Results: Averaging across the six models in group one led to a TTA of 69.3%, while the TTA of each single model ranged from 44.3% to 72.3%. MMA in group two exhibited a TTA of 70.8%, compared to a TTA of 44.3% to 66.6% with the six single models.

Conclusions: High variability in PP underscores the importance of model selection within MIPD for linezolid. The MAA achieved a TTA similar to or higher than the best performing single model without user intervention. The MAA is a promising alternative to the single-model approach, particularly in absence of external evaluation or if existing popPK models display unsatisfactory PP.

Keywords: Linezolid, MIPD, Pharmacokinetics, popPK, Model averaging

Background: Adherence to long-term pharmacotherapy plays a pivotal role in achieving optimal therapeutic outcomes. While therapeutic drug monitoring (TDM) is widely adopted for evaluating medication adherence, the current guidelines do not offer, or lack patient-specific reference range, potentially compromising the accuracy of adherence assessments.
Aims: The study aims to evaluate a pharmacokinetic-based Bayesian approach utilizing TDM measurements for evaluating medication adherence.
Methods: Taking neuropsychiatric drugs as an example for illustration, including anti-seizure medication, antipsychotic medication, anti-anxiety medication and others. Concentration-time profiles for virtual cohorts with monotherapy were simulated, including full adherence and diverse non-adherence patterns, based on population pharmacokinetic parameters derived from previous studies. Subsequently, a Bayesian framework was implemented to evaluate medication adherence based on TDM data. A systematic investigation of critical factors influencing adherence assessment was also conducted.
Results: By leveraging essential patient-specific data and dosing regimens, the integration of Bayesian approach with TDM data enables evaluation of recent medication adherence patterns. The concentration thresholds for adherence assessment demonstrated significant variability across different classes of neuropsychiatric medications, with notable influences from patients’ characteristics, concomitant pharmacotherapy, temporal sampling variations, dosing intervals, and prior probability distributions. To facilitate personalized adherence monitoring, an interactive web-based dashboard incorporating these critical variables was developed, which could be accessed via https://mipd.shinyapps.io/adherence_assessment/.
Conclusion: Pharmacokinetic-based Bayesian approach significantly enhances the capacity of TDM data in identifying distinct non-adherence patterns. These findings advocate for paradigm shift in therapeutic management toward personalized adherence assessment protocols.
Key words: anti-seizure medication, medication adherence, therapeutic drug monitoring, Bayesian method, population pharmacokinetics