External Validation and Bayesian Forecasting of Rivaroxaban Population Pharmacokinetic Models in Older Chinese Patients with Nonvalvular Atrial Fibrillation Authors: Jiale Chen, Weikun Huang, Guoquan Chen, Bangbiao Ge, Qiang Zhao, Haiping Wang, Jiahao Ye, Jianing Pan, Kaili Mao Published: 2026 ABSTRACT AbstractBackgroundPopulation pharmacokinetic (PopPK) models have been developed to characterize rivaroxaban exposure; however, their transferability across different clinical populations remains uncertain. This study aimed to externally evaluate published rivaroxaban PopPK models in older Chinese patients with nonvalvular atrial fibrillation (NVAF) and assess whether Bayesian forecasting could improve individual concentration prediction.MethodsSix published rivaroxaban PopPK models were evaluated using an independent cohort from the RIVA-GAP study. Predictive performance was assessed using prediction error metrics reflecting bias, precision, and prediction coverage, together with normalized prediction distribution error (NPDE) diagnostics. Bayesian forecasting was performed using paired steady-state trough–peak concentrations to evaluate changes in individual peak concentration prediction.ResultsA total of 135 patients with 257 rivaroxaban concentration observations were included for external validation, and 122 patients with paired trough–peak samples were included for Bayesian forecasting. Predictive performance varied substantially among the six models, and none consistently met predefined acceptance criteria. Model A showed comparatively lower prediction bias among the evaluated models but still demonstrated limited precision (MDPE 33.64%, MDAPE 53.72%, F30 31.52%). Bayesian forecasting reduced the MDIPE of Model A from 33.64% to 12.38%, but MAIPE remained 37.33% and IF30 was 41.8%. Similar model-dependent effects were observed across the other models.ConclusionPublished rivaroxaban PopPK models demonstrated heterogeneous transferability in older Chinese patients with NVAF. Bayesian forecasting using a single trough concentration provided modest and model-dependent improvement in concentration prediction but did not consistently overcome limitations of the underlying models.Clinical Trial/Study RegistrationChinese Clinical Trial Registry, ChiCTR2300074934; registered 21 August 2023.Keywords: rivaroxaban, atrial fibrillation, population pharmacokinetics, external validation, bayesian forecasting, older patients Introduction Atrial fibrillation (AF) is the most common cardiac arrhythmia, and is associated with a high risk of thromboembolic events, especially ischemic stroke.1 Effective long-term anticoagulation remains the cornerstone of stroke prevention in patients with AF. Direct oral anticoagulants (DOACs), including dabigatran and Factor Xa inhibitors such as rivaroxaban, are recommended by recent guidelines2,3 as first-line anticoagulants due to their predictable pharmacokinetics and pharmacodynamics (PK/PD), fixed-dose convenience, and favorable safety profile. Rivaroxaban is an oral, selective Factor Xa inhibitor, with a high bioavailability and dual elimination via hepatic metabolism and renal excretion.4 Despite the advantages of fixed dosing, substantial interindividual variability in rivaroxaban plasma concentrations persists, driven by patient characteristics, organ function, genetic polymorphisms, and concomitant medications.5 Population pharmacokinetic (PopPK) modeling offers a quantitative framework to address this variability.6,7 By analyzing concentration data using nonlinear mixed-effects methods, PopPK models characterize typical PK profiles, quantify interindividual variability, and identify significant covariates. Although several rivaroxaban PopPK models exist, few have been rigorously validated in independent cohorts. External validation is therefore essential to determine whether published rivaroxaban PopPK models can be reliably transferred to Chinese patients with NVAF. Accordingly, this study aimed (1) to externally evaluate the predictive performance and transferability of six published rivaroxaban PopPK models in this population, and (2) to explore whether incorporating a single prior trough concentration through Bayesian forecasting could improve prediction of a subsequent peak concentration. Materials and Methods Literature Search and Data Extraction A systematic literature search was conducted in major databases, including PubMed, Web of Science, Embase, and Cochrane Library, for PopPK or PK/PD studies of rivaroxaban published in English from its post-marketing period (September 16, 2008) to December 31, 2025. The detailed search strategy, including specific keywords and MeSH terms used for each database, is provided in Supplementary Table S1. Additional studies were identified by manually reviewing references of included articles. Studies were eligible if they: (1) employed nonlinear mixed-effects modeling (eg, NONMEM); (2) primarily involved patients with nonvalvular atrial fibrillation (NVAF) or included NVAF as a major subgroup; and (3) reported clear structural models, population PK parameters, inter-individual variability, and relevant covariates. Exclusion criteria included non-compartmental analyses, studies limited to healthy volunteers, incomplete data preventing extraction of key model parameters, and reviews, case reports, or conference abstracts. From each included study, information on study characteristics (author, year, design, sample size, age, sex, ethnicity) and model details (structure, estimation method, typical PK parameters, inter-individual and residual variability, covariates, and software) was extracted. NONMEM control streams were independently double-checked, and reconstructed Pop PK models were verified with corresponding authors when necessary. External Evaluation Dataset Rivaroxaban pharmacokinetic and demographic data were obtained from atrial fibrillation patients enrolled in the prospective RIVAroxaban Genetic Association in Patients with atrial Fibrillation (RIVA-GAP) study8 (ChiCTR2300074934), approved by the institutional review boards of participating hospitals. Written informed consent was obtained from all participants in accordance with the Declaration of Helsinki. Patients were considered eligible for external validation if they were aged ≥65 years and had available clinical characteristics, pharmacokinetic measurements, and covariate information required for application of the evaluated PopPK models, including body weight, lean body mass (LBM), renal function parameters, laboratory measurements, and ABCB1 rs4728709 genotype. For Bayesian forecasting analysis, only patients with paired steady-state trough and peak concentrations collected within the same dosing interval were included. The observed trough concentration was used as prior information for Bayesian estimation of individual pharmacokinetic parameters, which were subsequently used to predict the corresponding peak concentration. Rivaroxaban was administered as a fixed once-daily regimen according to routine clinical practice. Pharmacokinetic samples were collected under steady-state conditions after at least three consecutive days of continuous rivaroxaban administration. Trough samples were obtained immediately before the morning dose (approximately 0.5–1 hour before administration), and peak samples were collected during the expected absorption phase (approximately 2–4 hours after dosing). Plasma rivaroxaban concentrations were quantified using a validated liquid chromatography–tandem mass spectrometry (LC-MS/MS) method. The validated quantification range was 0.5–600 ng/mL. Detailed analytical procedures for LC-MS/MS measurement and ABCB1 rs4728709 genotyping are provided in the Supplementary Methods. Prediction-Based Diagnostics Predictive performance of the candidate models was assessed using observed concentrations (Cobs) and population-predicted concentrations (Cpre). The percentage prediction error (PE%), mean prediction error (MPE%), and root mean squared error (RMSE%) were calculated as follows: The median prediction error (MDPE) and median absolute prediction error (MDAPE) were employed to assess the accuracy and precision of the model prediction performance, respectively. Two additional metrics, the proportion of prediction error percentage falling within the ±20% range (F20) and the proportion within the ±30% range (F30), were calculated as combined indicators of accuracy and precision. The predefined criteria for acceptable predictive performance were an absolute value of MDPE ≤ 20%, MDAPE ≤ 30%, F20 ≥ 35%, and F30 ≥ 50%.9 Because external applicability reflects multiple aspects of predictive performance, model interpretation was based on the overall pattern of bias, precision, and coverage metrics rather than a single statistical indicator. Simulation-Based Diagnostics Normalized prediction distribution errors (NPDE) were computed separately for each evaluated PopPK model based on 2000 simulations generated from the corresponding NONMEM model. NPDEs were expected to follow a standard normal distribution, assessed using the Wilcoxon signed-rank test (mean = 0), Fisher variance test (variance = 1), Shapiro–Wilk test (normality), and a Global test (minimum p-value multiplied by 3).10 Because formal normality tests may be sensitive to relatively moderate sample sizes and minor deviations, NPDE interpretation was complemented by graphical diagnostics, including NPDE distribution plots, NPDE versus population-predicted concentration plots, and normal quantile–quantile (Q-Q) plots. These graphical evaluations were used together with statistical tests to assess systematic bias, dispersion, and deviations from normality. Bayesian Forecasting Bayesian forecasting was performed to evaluate whether incorporation of a single observed trough concentration could improve individual prediction of rivaroxaban peak concentrations. For patients with paired steady-state trough–peak samples, the observed trough concentration was incorporated as prior information to estimate individual pharmacokinetic parameters using Bayesian maximum a posteriori estimation. The estimated individual parameters were subsequently used to predict the peak concentration within the same dosing interval. Individual prediction error (IPE%) was calculated as the relative difference between Bayesian-predicted peak concentration and observed peak concentration. Predictive performance after Bayesian updating was evaluated using median individual prediction error (MDIPE%), median absolute individual prediction error (MAIPE%), and the proportions of predictions within ±20% and ±30% error ranges (IF20 and IF30). Changes in prediction error distributions before and after Bayesian forecasting were compared across the six evaluated PopPK models. Statistical Analysis Continuous variables were summarized as median (interquartile range) or mean (standard deviation), as appropriate. Categorical variables were presented as numbers and percentages. All statistical analyses were performed using R software and NONMEM. Statistical tests were interpreted descriptively, and a two-sided P value <0.05 was considered statistically significant where applicable. This study was reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement.11 Results Characteristics of External Dataset A total of 222 patients enrolled in the RIVA-GAP study were screened, and 135 patients with complete demographic, pharmacokinetic, genetic, and covariate information were included in the external validation cohort (Figure 1). After exclusion of six concentration observations outside the validated LC-MS/MS quantification range (0.5–600 ng/mL), 257 rivaroxaban concentration observations from 135 patients were available for external validation. Among them, 122 patients with paired steady-state trough and peak concentrations were included in the Bayesian forecasting analysis (Figure 1). Flow diagram of cohort selection and pharmacokinetic analysis populations Patients enrolled in the RIVA-GAP study were sequentially screened according to predefined eligibility criteria. After exclusion of patients with missing demographic, genetic, or pharmacokinetic information, a complete-case cohort was established for external validation of six published rivaroxaban population pharmacokinetic models. Concentration observations outside the validated LC-MS/MS quantification range were excluded before pharmacometric evaluation. Patients with paired trough–peak concentration samples were further included for Bayesian forecasting analysis. The baseline characteristics of the external validation cohort are summarized in Table 1. The median age was 75.0 years (IQR 70.0–81.0), and 65 patients (48.1%) were male. The median body weight and lean body mass were 59.5 kg (IQR 52.3–66.0) and 43.3 kg (IQR 39.7–50.6), respectively. Most patients received rivaroxaban 10 mg once daily (58.5%), followed by 15 mg once daily (37.0%). The ABCB1 rs4728709 dominant genotype (AA/GA) was present in 41 patients (30.4%). Laboratory parameters included alanine aminotransferase of 21.0 U/L (IQR 14.0–33.0), hematocrit of 39.5% (SD 5.6), total bilirubin of 17.2 μmol/L (IQR 13.2–22.9), blood urea nitrogen of 20.1 mg/dL (IQR 15.7–25.7), and serum creatinine of 92.0 μmol/L (IQR 75.9–110.0). The most common comorbidities were hypertension (67.4%), heart failure (63.0%), and type 2 diabetes mellitus (24.4%). Table 1Demographic and Clinical Characteristics of the External Validation PatientsVariableAvailable nOverall (N = 135)Personal characteristicsSex, male13565 (48.1%)Age, years13575.0 (70.0–81.0)Body weight, kg13559.5 (52.3–66.0)Lean body mass, kg13543.3 (39.7–50.6)Rivaroxaban daily dose135 5 mg2 (1.5%) 10 mg79 (58.5%) 15 mg50 (37.0%) 20 mg4 (3.0%)ABCB1 rs4728709 genotype (AA/GA)13541 (30.4%)Laboratory testsAlanine aminotransferase, U/L13521.0 (14.0–33.0)Hematocrit, %13539.5 (5.6)Total bilirubin, μmol/L13517.2 (13.2–22.9)Blood urea nitrogen, mg/dL13520.1 (15.7–25.7)Serum creatinine, umol/L13592.0 (75.9–110.0)ComorbiditiesHypertension13591 (67.4%)Type 2 diabetes mellitus13533 (24.4%)Malignancy1355 (3.7%)Heart failure13585 (63.0%)Left ventricular dysfunction1359 (6.7%)Coronary heart disease13566 (48.9%)Chronic obstructive pulmonary disease1356 (4.4%)Concomitant medicationsAntiplatelet drugs13517 (12.6%)Non-steroidal anti-inflammatory drugs1354 (3.0%)Proton pump inhibitors13544 (32.6%)Statins13588 (65.2%)Angiotensin receptor blockers13561 (45.2%)Beta-blockers13580 (59.3%)Calcium channel blockers13519 (14.1%)Digoxin13544 (32.6%)Notes: Values are presented as n (%), median (interquartile range), or mean (standard deviation), as appropriate.Abbreviations: ABCB1, ATP-binding cassette subfamily B member 1; NSAIDs, non-steroidal anti-inflammatory drugs; PPIs, proton pump inhibitors; ARB, angiotensin receptor blockers; COPD, chronic obstructive pulmonary disease. Demographic and Clinical Characteristics of the External Validation Patients Notes: Values are presented as n (%), median (interquartile range), or mean (standard deviation), as appropriate. Abbreviations: ABCB1, ATP-binding cassette subfamily B member 1; NSAIDs, non-steroidal anti-inflammatory drugs; PPIs, proton pump inhibitors; ARB, angiotensin receptor blockers; COPD, chronic obstructive pulmonary disease. Characteristics of Evaluated PopPK Models Of 308 records initially identified, 107 duplicates were removed, leaving 201 records for screening. Full-text assessment excluded 38 studies due to incomplete data or non-use of NONMEM modeling, resulting in six studies included for external validation (Supplementary Figure S1). The six published rivaroxaban PopPK models12–17 differed in compartmental structure, key covariates, software platforms, and internal evaluation methods (Table 2). All models were developed using nonlinear mixed-effects modeling approaches, with one-compartment structures used across the evaluated models. Detailed characteristics of the included PopPK models are presented in Supplementary Table S2. Covariates incorporated across these models included body weight, lean body mass, age, renal function (creatinine clearance or estimated glomerular filtration rate), total bilirubin, blood urea nitrogen, hematocrit, and ABCB1 genotype, and concomitant medication use. Internal validation reported in the original publications was conducted using goodness-of-fit plots, visual predictive checks, and bootstrap analyses. External evaluation in independent populations was limited. Table 2Summary of the Evaluated Population Pharmacokinetic ModelsModel SourceCompartmentsKey CovariatesSoftwarePharmacokinetic ParameterInternal EvaluationExternal Evaluation in Original PublicationModel ApplicationModel AZhang et al,12OneV: body weight; CL: creatinine clearance, total bilirubin, and ABCB1 rs4728709 genotypeNONMEM 7.4GOF plots, VPC, and bootstrap analysisNot reportedSimulated steady-state exposure after 10, 15, and 20 mg once daily across covariate levels including CrCL, total bilirubin, genotype, and body weight; exposures were compared with target concentration rangesModel BGirgis et al,13OneV: lean body mass and age; CL: age and serum creatinineNONMEM V level 1.1 and 7.1.0; S-PLUS 6.2; RPrediction-corrected VPC, GOF plots, and CWRES distribution plotsNot reportedCompared exposure in patients with moderate renal impairment receiving 15 mg with exposure in patients with normal or mild renal impairment receiving 20 mgModel CLiu et al,14OneCL: eGFR; PT: eGFR and total bilirubinNONMEM 7.4.2GOF plots, bootstrap analysis (n = 2,000), and VPC (n = 1,000)Not reportedUsed Monte Carlo simulation to compare exposure across ethnic groups and renal function categoriesModel DTanigawa et al,15OneCL: blood urea nitrogenNONMEM version V level 1.1GOF plots, VPC, and bootstrap analysis (n = 1,000)Not reportedSimulated steady-state exposure (AUCss and Cmax,ss) after 15 and 20 mg in Japanese patients and compared these results with exposure after 20 mg in Caucasian patientsModel EWillmann et al,16OneV: body weight, age, and sex;CL: CrCL, body weight, concomitant medication, and study population;F:doseNONMEM 7.3GOF plots, CWRES plots, pcVPC, and bootstrap analysis (n = 200)Not reportedSimulated steady-state exposure (AUC0-24, Cmax, and Ctrough) across age, renal function, and body weight/BMI subgroupsModel FSuzuki et al,17OneCL: CrCL, ALT, and concomitant CYP3A4 or P-gp inhibitors;PT: hematocritPhoenix NLME 1.4GOF plots, CWRES plots, and prediction-corrected VPCNot reportedReported similar exposure in Japanese patients receiving 15 mg and Caucasian patients receiving 20 mgNotes: External evaluation refers to evaluation reported in the original model publication, not the external validation performed in the present study. For ABCB1 rs4728709, rs4728709 = 1 indicates AA or GA genotype; rs4728709 = 0 indicates GG genotype.Abbreviations: ALT, alanine aminotransferase; AUC, area under the concentration-time curve; BMI, body mass index; CL, clearance; Cmax, maximum concentration; Ctrough, trough concentration; CrCL, creatinine clearance; CWRES, conditional weighted residuals; eGFR, estimated glomerular filtration rate; GOF, goodness of fit; pcVPC, prediction-corrected visual predictive check; PT, prothrombin time; V, volume of distribution; VPC, visual predictive check. Summary of the Evaluated Population Pharmacokinetic Models Notes: External evaluation refers to evaluation reported in the original model publication, not the external validation performed in the present study. For ABCB1 rs4728709, rs4728709 = 1 indicates AA or GA genotype; rs4728709 = 0 indicates GG genotype. Abbreviations: ALT, alanine aminotransferase; AUC, area under the concentration-time curve; BMI, body mass index; CL, clearance; Cmax, maximum concentration; Ctrough, trough concentration; CrCL, creatinine clearance; CWRES, conditional weighted residuals; eGFR, estimated glomerular filtration rate; GOF, goodness of fit; pcVPC, prediction-corrected visual predictive check; PT, prothrombin time; V, volume of distribution; VPC, visual predictive check. Prediction-Based External Validation Prediction-based diagnostics demonstrated considerable variability in predictive performance among the six evaluated rivaroxaban PopPK models (Figure 2 and Supplementary Table S3). The distribution of individual prediction errors showed substantial interindividual variability across all models. Models A, B, D, and E demonstrated prediction errors with medians closer to zero compared with Models C and F, whereas Model F showed a marked tendency toward overprediction, reflected by a MDPE of 130.53%. External validation performance of six published rivaroxaban population pharmacokinetic models based on population predictions. Box plots illustrate the distribution of individual prediction errors (PE, %) for six previously published rivaroxaban population pharmacokinetic models applied to the external validation dataset. Each point represents an individual concentration observation. The solid horizontal line indicates zero prediction error, and dashed lines represent ±20% and ±30% prediction error boundaries. Prediction errors were calculated by comparing model-predicted concentrations with corresponding observed rivaroxaban concentrations from all evaluable observations within the validated LC-MS/MS quantification range. Quantitative evaluation showed that Model A had the lowest prediction bias among the evaluated models, with an MPE of 135.1% and MDPE of 33.64%. However, the prediction precision remained limited, with an MDAPE of 53.72%, F20 of 21.01%, and F30 of 31.52%. Models B–E showed comparable prediction performance, with MDPE values ranging from 33.64% to 63.47% and F30 values ranging from 27.24% to 31.52%. Model F demonstrated the poorest performance, with the highest MDPE (130.53%), MDAPE (153.47%), and lowest F20 (11.67%) and F30 (18.29%). Overall, none of the evaluated models simultaneously fulfilled all predefined acceptance criteria, indicating limited direct transferability of currently available rivaroxaban PopPK models to this cohort of older Chinese patients with NVAF. NPDE-Based External Validation Diagnostics NPDE analyses demonstrated model-dependent differences in prediction behavior among the six evaluated PopPK models (Table 3). NPDE diagnostics demonstrated model-dependent deviations. Model A showed relatively balanced dispersion (variance 1.127), although residual bias remained (Wilcoxon P < 0.001), indicating incomplete agreement between predicted and observed distributions. Models B–E exhibited varying degrees of location bias or dispersion abnormalities, whereas Model F showed the largest deviation, with a negative NPDE shift (median −0.992) and the highest proportion of observations outside ±1.96 (18.3%). Table 3NPDE-Based External Validation Diagnostics of Six Published Rivaroxaban Population Pharmacokinetic ModelsModelnMeanNPDESDVarianceMedian (IQR)|NPDE| > 1.96,n (%)NPDE >1.96/NPDE<-1.96, n/nWilcoxon PVariancePShapiro–Wilk PGlobal PModel A257−0.2931.0621.127−0.264 (−0.988, 0.323)25 (9.7)5/203.03E-05*0.1580.1159.09E-05*Model B257−0.2860.9400.884−0.395 (−0.987, 0.313)18 (7.0)5/132.13E-07*0.1790.00295*6.39E-07*Model C257−0.4711.1471.315−0.345 (−1.145, 0.458)22 (8.6)5/173.67E-10*0.00105*0.00868*1.10E-09*Model D257−0.4961.0041.008−0.645 (−1.159, 0.212)27 (10.5)6/217.45E-13*0.9050.00414*2.24E-12*Model E257−0.3810.9240.854−0.440 (−1.026, 0.195)15 (5.8)4/111.02E-09*0.0870.1313.05E-09*Model F257−0.9330.9190.845−0.992 (−1.621, −0.252)47 (18.3)3/441.05E-30*0.06785.28E-04*3.16E-30*Abbreviations: NPDE, normalized prediction distribution error; IQR, interquartile range. Wilcoxon signed-rank test assessed whether the median NPDE differed from zero; the variance test assessed whether NPDE variance differed from one; the Shapiro–Wilk test assessed normality. Global P values were calculated using a Bonferroni correction by multiplying the minimum P value among the three diagnostic tests by three, capped at 1. *P < 0.05. NPDE-Based External Validation Diagnostics of Six Published Rivaroxaban Population Pharmacokinetic Models Abbreviations: NPDE, normalized prediction distribution error; IQR, interquartile range. Wilcoxon signed-rank test assessed whether the median NPDE differed from zero; the variance test assessed whether NPDE variance differed from one; the Shapiro–Wilk test assessed normality. Global P values were calculated using a Bonferroni correction by multiplying the minimum P value among the three diagnostic tests by three, capped at 1. *P < 0.05. Graphical NPDE diagnostics, including NPDE distribution plots, NPDE versus population-predicted concentration plots, and normal Q-Q plots, were evaluated together with statistical tests (Supplementary Figure S2). These graphical assessments further supported the presence of model-specific bias and dispersion abnormalities, indicating that none of the evaluated models completely reproduced the observed concentration distribution in the external validation cohort. Bayesian Forecasting Bayesian forecasting was performed in 122 patients with paired steady-state trough and peak rivaroxaban concentrations. Compared with population-based predictions, incorporation of a single observed trough concentration resulted in variable changes in individual prediction error distributions across the six evaluated models (Figure 3 and Supplementary Table S4). For Models A, B, D, and E, Bayesian updating narrowed the distribution of individual prediction errors and shifted prediction error medians toward zero, suggesting reduced prediction variability. Comparison of individual prediction errors before and after Bayesian forecasting using prior trough concentrations. Box plots compare individual prediction errors before and after Bayesian forecasting for the six evaluated rivaroxaban population pharmacokinetic models. Bayesian forecasting incorporated observed trough concentrations as prior information to estimate individual pharmacokinetic parameters and predict peak concentrations. Each point represents an individual prediction error. The solid horizontal line indicates zero prediction error, and dashed lines represent ±20% and ±30% prediction error boundaries. For Model A, Bayesian forecasting reduced the MDIPE from 33.64% to 12.38% and increased IF20 from 21.01% to 35.25%. However, the median absolute prediction error remained above the predefined criterion (MAIPE: 37.33%), and IF30 remained below the acceptance threshold (41.8%). Models B and E showed reductions in median prediction bias after Bayesian updating, with MDIPE values changing to −3.58% and −3.61%, respectively. Nevertheless, improvements in overall prediction precision were limited, as IF20 and IF30 values did not consistently reach predefined performance criteria. In contrast, Model F showed minimal improvement after Bayesian updating, with persistent prediction bias (MDIPE: 135.95%) and low prediction coverage (IF20: 9.84%, IF30: 18.03%). Discussion In this study, we performed an external evaluation of six published rivaroxaban PopPK models in an independent cohort of older Chinese patients with NVAF. Our findings demonstrate substantial heterogeneity in model transferability across different patient populations. Although several models showed relatively lower prediction bias, none simultaneously fulfilled all predefined external validation criteria, highlighting that published PopPK models should not be considered directly interchangeable without evaluation in the intended clinical population. Furthermore, Bayesian forecasting incorporating a single prior trough concentration resulted in modest and model-dependent changes in individual peak concentration prediction. While prediction bias was reduced for several models, improvements in overall prediction precision remained limited. The heterogeneous performance of the evaluated models likely reflects differences in model structure, covariate selection, and characteristics of the development populations. Model A12 showed relatively favorable external performance, which may partly relate to the inclusion of clinically relevant covariates.13,16 However, covariate inclusion alone does not guarantee successful model transferability. Differences in demographic characteristics, disease status, and covariate distributions between development and validation cohorts may contribute to reduced predictive performance. For example, Model C14 incorporated estimated glomerular filtration rate-based covariates, and differences in renal function assessment18 and patient characteristics may partly contribute to reduced applicability. Models B13 and E16 were developed in populations that differed from the present elderly Chinese NVAF cohort, and ethnic-related pharmacokinetic differences may partially influence model applicability. The apparent clearance of rivaroxaban has been reported to be approximately 30–40% lower in Asian adults than in Caucasian adults.19 Similarly, the poorer performance of Model F17 may reflect differences in development population and model structure, including the absence of body size-related covariates. Although ABCB1 rs4728709 may contribute to interindividual variability, genotype distributions across previous model development cohorts were insufficiently reported, and its contribution to model transferability remains uncertain. Bayesian forecasting20 has been increasingly applied in pharmacokinetic modeling to integrate population information with patient-specific drug concentrations and refine individual exposure estimation. In the present study, incorporation of a single observed trough concentration reduced prediction bias for several rivaroxaban PopPK models, suggesting that limited pharmacokinetic sampling may provide additional information beyond population-based predictions. However, the magnitude of improvement varied across models, and reductions in bias did not consistently translate into sufficient improvement in prediction precision. For example, although Bayesian updating improved the median individual prediction error of Model A, the remaining absolute prediction error and prediction coverage did not fully meet predefined criteria. These findings indicate that Bayesian forecasting cannot completely compensate for limitations in the underlying PopPK model, and the accuracy of individual prediction remains dependent on the validity and transferability of the selected model.21 From a clinical perspective, our findings emphasize that external validation is an essential step before applying PopPK models to specific patient populations. Although Bayesian forecasting represents a promising approach for improving exposure estimation when limited sampling information is available, its clinical implementation requires consideration of model suitability, analytical reliability, and prospective evaluation. Importantly, improved pharmacokinetic prediction does not necessarily translate into improved clinical outcomes, and further studies are needed to determine whether Bayesian-guided exposure assessment can contribute to improved anticoagulation exposure assessment. Several limitations should be acknowledged. First, although this study provided an independent external evaluation of published rivaroxaban PopPK models, the validation cohort consisted of older Chinese patients with NVAF, and the generalizability of our findings to other populations requires further investigation. Second, Bayesian forecasting was limited to patients with paired trough–peak concentrations, and clinical practice-based sampling may not fully characterize individual pharmacokinetic variability. Third, this study focused on pharmacokinetic prediction performance rather than clinical outcomes; whether improved exposure prediction translates into better anticoagulation management remains to be determined. Finally, residual variability may remain due to unmeasured factors, including adherence, concomitant medications, and additional genetic determinants. Conclusion In conclusion, published rivaroxaban PopPK models showed heterogeneous transferability in older Chinese patients with NVAF, and none consistently met predefined external validation criteria. Bayesian forecasting using a single trough concentration provided modest and model-dependent improvement in concentration prediction but could not fully overcome limitations of the underlying models. Further studies incorporating pharmacodynamic markers and clinical outcomes are needed to determine whether improved exposure prediction translates into clinically meaningful benefits. Acknowledgments The authors thank all participants and clinical staff at the participating centers for their support of this study. Funding Statement This work was supported by the Jinhua Central Hospital Youth Research Foundation (Grant Nos. JY2022-2-03 and JY2024-2-14), the Health Project of Longyan Joint Fund for Scientific and Technological Innovation (Grant No. 2023LYF17089), and Key Science and Technology Plan Projects of Jinhua (Grant No: 2023-3-122). The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data Sharing Statement The datasets generated and/or analyzed during the current study are not publicly available due to privacy and ethical restrictions. De-identified data may be made available from the corresponding author upon reasonable request and subject to institutional and ethical approval. Code Availability The code and model files used during the current study are not publicly available. They may be made available from the corresponding author upon reasonable request and subject to appropriate approval. Ethics Approval This study was approved by the Ethics Committee of Affiliated Jinhua Hospital, Zhejiang University School of Medicine (Approval No. [2023]-Ethics Approval No. 05) and the Ethics Committee of Longyan Second Hospital (Approval No. LYEYKY-2024-006). The study was conducted in accordance with the principles of the Declaration of Helsinki. Consent to Participate Written informed consent was obtained from all participants included in the study. Clinical Trial/Study Registration This study was registered in the Chinese Clinical Trial Registry (ChiCTR2300074934; registration date: 21 August 2023). Author Contributions All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Disclosure The authors declare that they have no competing interests relevant to the content of this article. References Associated Data Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available due to privacy and ethical restrictions. De-identified data may be made available from the corresponding author upon reasonable request and subject to institutional and ethical approval.