Development and Clinical Application of a Real-World Population Pharmacokinetic Model of Rivaroxaban in Asian Patients with Atrial Fibrillation Authors: Nian-Shiuan Lai, Chun-Jung Lin, Ching-Hua Kuo, Yu-Fong Peng, Sung-Chun Tang, Chih-Fen Huang, Shin-Yi Lin, Shu-Wen Lin Published: 2026 ABSTRACT AbstractBackgroundRivaroxaban, a factor Xa inhibitor, is used to prevent stroke in patients of atrial fibrillation (AF). This study aimed to establish a population pharmacokinetic (PPK) model for rivaroxaban using real-world data, apply it to concentration prediction, and investigate the association between drug exposure and clinical outcomes.MethodsPatients with AF receiving rivaroxaban were enrolled from an observational cohort (2016–2023) to measure plasma concentrations. An independent cohort was randomly selected to validate the PPK model. Clinical outcomes of interest included stroke or systemic thromboembolism, and major bleeding.ResultsA total of 226 patients contributed to 452 rivaroxaban concentration measurements. Rivaroxaban pharmacokinetics were adequately described using a one-compartment model with first-order elimination. The estimated apparent clearance (CL/F) and the volume of distribution (V/F) were 6.13 L/h and 45.57 L, respectively. CL/F was significantly influenced by creatinine clearance and concomitant use of cytochrome 3A4 or P-glycoprotein inhibitors, whereas V/F was associated with lean body weight. External validation demonstrated a good predictive performance at the individual level. Patients with low trough concentrations tended to have an increased risk of systemic thromboembolism, whereas those with high trough concentrations tended to have a higher risk of major bleeding.ConclusionsIn an Asian population, rivaroxaban pharmacokinetics are influenced by renal function, lean body weight, and drug interactions. The developed PPK model facilitates the estimation of rivaroxaban concentrations at standardized timepoints from random samples. This provides a practical tool for standardized exposure assessment and the identification of patients at risk for adverse clinical outcomes.RegistrationClinicalTrials.gov identifier no. NCT05333666.Supplementary InformationThe online version contains supplementary material available at 10.1007/s40262-026-01650-4. Key Points On the basis of real-world data, rivaroxaban levels were affected by lean bodyweight, renal function, and drug interactions.The PPK model facilitates the estimation of rivaroxaban concentrations at standardized time points from random samples, supporting more personalized anticoagulant treatment. Introduction Rivaroxaban is a factor Xa inhibitor and the first-line therapy recommended for the prevention of stroke and systemic thromboembolism (SSE) in patients with atrial fibrillation (AF) [1]. In both the Rivaroxaban Once-Daily Oral Direct Factor Xa Inhibition Compared with Vitamin K Antagonism for Prevention of Stroke and Embolism Trial in Atrial Fibrillation (ROCKET AF) trial and the J-ROCKET AF study conducted in Japan, rivaroxaban was shown to be noninferior to warfarin in preventing SSE, with a comparable risk of major bleeding and a lower risk of intracranial hemorrhage (ICH) [2, 3]. Rivaroxaban exhibits linear pharmacokinetics at doses up to 15 mg. However, at a higher dose of 15 or 20 mg, the pharmacokinetics become nonlinear due to solubility-limited absorption. While the oral bioavailability remains high (80–100%) for 10 mg doses, it decreases to approximately 66% for the 15 or 20 mg dose when administered in a fasting state. Therefore, the co-administration of food is recommended. Approximately 36% of the administered rivaroxaban dose is excreted unchanged in urine. The remainder undergoes metabolic degradation, including 18% by cytochrome P (CYP)3A4, 14% by CYP2J2, and 14% by non-CYP-mediated hydrolysis via renal or hepatobiliary routes. The half-life of rivaroxaban is 5–9 h in healthy individuals and is prolonged in elderly patients or those with renal impairment patients [4]. Population pharmacokinetic (PPK) modeling analyzes drug concentrations and variability within a patient population, and characterizes the factors associated with interindividual variability in drug exposure [5]. The PPK properties of rivaroxaban have been characterized in several studies using clinical trial data. For example, phase II study data were used in PPK analysis to support dose selection in a phase III trial and to compare rivaroxaban exposure between Japanese patients and those in the global ROCKET-AF study [6, 7]. Additionally, an integrated PPK model was established using pooled data from seven randomized controlled trials across various rivaroxaban indications to predict rivaroxaban exposure in patient subgroups of special interests [8]. In real-world practice, patients often present with complex comorbidities and physiological conditions that differ significantly from those in strictly controlled clinical trials. Furthermore, dosing patterns in routine care may deviate from standardized trial protocols [9, 10]. While several PPK models for rivaroxaban have been established using clinical trial data [8, 11], their applicability to real-world populations with a wider range of renal function and lower body weights remains insufficiently characterized. Although some real-world PPK models have been reported [12–14], they are often limited by small sample sizes and a lack of correlation with clinical outcomes. Furthermore, a defined therapeutic range for rivaroxaban based on these models is currently lacking. Therefore, this study aimed to establish and externally validate a PPK model of rivaroxaban using data from a prospectively enrolled observational cohort. Most importantly, we sought to identify real-world covariates and evaluate the association between model estimated concentrations and the risk of clinical outcomes, including SSE and major bleeding, to assess the model’s utility in supporting individualized anticoagulation management. Methods Participant Data of this study were collected from the National Taiwan University Hospital (NTUH), a tertiary medical center in northern Taiwan. The study population was derived from a DOAC-T (direct oral anticoagulant-Taiwan) cohort that prospectively included patients with AF receiving direct oral anticoagulant (DOAC) therapy. Patients were eligible if they were > 20 years old and had been receiving rivaroxaban therapy for at least 3 days. Patients were excluded if they had end-stage renal disease requiring dialysis, failed to provide at least two blood samples, missed the rivaroxaban dose on the day before blood sampling, were lost to follow-up at the main and branch hospitals of NTUH within 14 days after blood collection, or declined to provide informed consent. The study protocol was approved by the International Ethics Committee of NTUH (no. 202201014RINB) and registered in the ClinicalTrials.gov database (NCT05333666). Study Procedure For all participants, two blood samples were collected at steady-state (defined as ≥ 3 days of continuous rivaroxaban use): one at peak concentration (2–4 h after administration), and one at trough (immediately before the next scheduled dose). The peak sample was collected following a supervised dose administered by a pharmacist. Trough samples reflected variations in dosing intervals in clinical practice. The actual sampling times were recorded as time-since-last-dose to capture the dosing variations. Plasma rivaroxaban concentrations were measured using validated ultra-high-performance liquid chromatography coupled with tandem mass spectrometry (UHPLC-MS/MS) [15]. Details are provided in the Supplementary Materials. Demographics, comorbidities, laboratory results, and concurrent medications were retrieved from the electronic medical records (EMR). Laboratory parameters included blood urea nitrogen (BUN), serum creatinine (CRE), aspartate aminotransferase, alanine aminotransferase, total bilirubin, hematocrit, and serum albumin. Concurrent medications were classified into four categories: cytochrome P450 3A4 (CYP3A4) or P-glycoprotein (P-gp) inhibitors, CYP3A4 or P-gp inducers, cytochrome P450 2J2 (CYP2J2) inhibitors, and breast cancer resistance protein (BCRP) inhibitors. A detailed list is provided in Table S1. To estimate creatinine clearance, the Cockcroft–Gault equation was utilized, while the estimated glomerular filtration rate was calculated using the four-variable modification of diet in renal disease (MDRD-4) equation. To estimate fat-free body mass, we evaluated two descriptors: the Janmahasatian formula (lean body weight, LBW) and the James formula (lean body mass, LBM) [16, 17]. The former is considered the contemporary standard in PPK studies, whereas the latter is a traditional body-size descriptor. The detailed equations are provided in Table S2. PPK Model Establishment The PPK model was developed using Monolix version 2023R1 (Lixoft SAS, Antony, France). Model development was performed using a nonlinear mixed-effects modeling approach, beginning with a base model that included a structural model and random effects, followed by covariate analysis to identify significant predictors of pharmacokinetic variability, as presented in Fig. S1. On the basis of previous investigations in patients with AF receiving rivaroxaban, a one-compartment model with first-order absorption and elimination was utilized [6, 8, 11]. This structural choice was appropriate for the sparse sampling design of our cohort and was further evaluated through visual inspection of the observed concentration–time profiles. Given that only one post-dose peak sample was collected and absorption kinetics may not be adequately characterized, the absorption rate constant (Ka) was fixed at 0.821 h−1 according to literature findings [8]. While the structural framework was based on previous investigations, all primary pharmacokinetic parameters and variability components were re-estimated using our real-world Asian cohort to capture the specific characteristics of this population. Both additive and exponential random effects models were used to evaluate interindividual variability (IIV). The residual variability (RV) was examined using proportional, additive, and combined error models. Model selection was based on diagnostic plots, visual predictive checks, and changes in objective function values (OFV). Covariates were selected using a stepwise forward inclusion and backward elimination approach (Fig. S1). Covariates were individually added to the base model, and a decrease in OFV > 3.84 (P < 0.05) was considered statistically significant for inclusion. Following the construction of the full model, covariates were removed sequentially and retained only if their removal caused an OFV increase of > 6.63 (P < 0.01) relative to the full model. This more stringent criterion during elimination ensures that the final model is parsimonious and includes only the most robust predictors of pharmacokinetic variability. Model stability and parameter precision were evaluated using relative standard error (RSE), with RSE < 30% considered acceptable for fixed effects and < 40–50% for random effects. The coefficient of variation (CV) was used to quantify IIV magnitude. Goodness-of-fit (GOF) was assessed using scatter plots of observed versus predicted concentrations (population prediction [PRED] and individual predictions [IPRED]) and plots of individual weighted residuals (IWRES) against predictions. To verify the model’s predictive capability, a visual predictive check (VPC) was performed on the basis of 1000 simulations using the original dataset. The 5th, 50th, and 95th percentiles of the observed data were overlaid with the 90% prediction intervals of the simulated data to ensure the model adequately captured the central tendency and variability of rivaroxaban concentrations. Notably, extreme concentration values were excluded from the PPK model establishment on the basis of the following criteria: (1) analytical outliers, defined as concentrations below the limit of quantification; and (2) a statistical outlier. Validation Cohort To validate the PPK model, an independent cohort of patients with AF receiving rivaroxaban was randomly selected from the DOAC-T database, excluding individuals used for model establishment. Predicted rivaroxaban concentrations were calculated using patient-specific covariates and compared with the observed concentrations to assess predictive accuracy. Model Prediction of Drug Concentration The final PPK model was applied using Monolix (version 2023R1; Lixoft SAS, Antony, France) to estimate the individual plasma concentrations at 3 h post-dose (peak concentration, Cpk) and 24 h post-dose (trough concentration, Ct) for each participant in the PPK modeling cohort. The area under the concentration–time curve (AUC) was estimated as the administered dose divided by apparent clearance (CL/F), with units of ng × h/mL. Participants were categorized into high- or low-exposure groups on the basis of the expected concentration and AUC ranges reported in clinical trials. These ranges, representing the 5th to 95th percentiles of the concentration distribution, were 178–343 ng/mL for Cpk, 12–137 ng/mL for Ct, and 1860–5434 ng × h/mL for AUC [6, 18]. Specifically, high exposure was defined as estimated Cpk, Ct, or AUC value exceeding the upper bound of the reference ranges. Conversely, low exposure was defined as any value falling below the corresponding lower bound. The incidence of clinical events was compared across both concentration- and AUC-defined groups to determine which metric correlated with the clinical outcomes. Clinical Outcomes Clinical outcomes were prospectively collected from the date of rivaroxaban concentration measurement until the occurrence of a clinical event, death, loss to follow-up at the main or branch hospitals of NTUH, or the study end date (31 August 2024), whichever occurred first. Outcomes of interest included SSE and major bleeding. SSE included ischemic stroke, transient ischemic attack (TIA), and systemic thromboembolism. Major bleeding was defined according to the platelet inhibition and patient outcomes (PLATO) criteria [19]. Major life-threatening bleeding was defined as ICH, intrapericardial bleeding, overt bleeding that led to hemorrhagic shock, hypotension requiring inotropic agents or surgical intervention, or a reduction in hemoglobin of more than 5 g/dL or requiring packed red blood cell transfusion of ≥ 4U. Other major bleeding was defined as significant, disabling bleeding or bleeding that resulted in a drop in the hemoglobin level by 3–5 g/dL, thereby requiring a packed red blood cell transfusion of 2–3 U. Statistical Analysis Data are presented as mean ± standard deviation for continuous variables and as frequencies (%) for categorical variables. Group comparisons of baseline characteristics were conducted using Student’s t-test or one-way analysis of variance (ANOVA) for continuous data, and the chi-squared test for categorical data. During PPK model development, Pearson’s correlation was used to assess the relationship between continuous clinical variables and pharmacokinetic parameters, while Student’s t-test or ANOVA was applied to evaluate categorical covariates. To evaluate the association between rivaroxaban exposure and clinical outcomes, we utilized Kaplan–Meier survival curves with the log-rank test to visualize and compare the cumulative incidence of SSE and major bleeding across exposure strata. To quantify the impact of drug exposure on event risk, Cox proportional hazards models were employed to estimate hazard ratios (HRs) and 95% confidence intervals (CIs). Initial analyses treated rivaroxaban concentrations and AUC values as continuous variables to assess exposure response trends. These were followed by categorical analyses on the basis of expected ranges reported in clinical trials, a method previously used in our research to classify patients at risk of clinical events [10, 20]. All analyses were two-tailed, with P < 0.05 considered statistically significant, and were performed using R software (version 4.3.1). Sensitivity Analysis To minimize potential distortion of parameter estimates, the initial PPK model was developed after excluding extreme values. However, as these values might represent true concentrations, their exclusion could result in the loss of important information. Therefore, in the first sensitivity analysis, we re-estimated the model including all values. Second, to evaluate whether the model was sensitive to assumptions regarding the absorption phase, we reestablished the model using an alternative Ka of 0.671 h−1. This independent analysis verified the stability of our findings across different potential absorption rates and confirmed the robustness of our conclusions [6]. Lastly, due to a high proportion of missing values for several laboratory parameters, only CRE and creatinine clearance (CrCL) were included as candidates for covariate screening during the initial PPK modeling. To ensure that no influential predictors were overlooked, we performed a diagnostic assessment of the correlation between the CL/F and the variables with missing data. This evaluation was conducted to determine whether any of these parameters exhibited a significant trend that would warrant their incorporation into the final PPK model. Results Basic Characteristics A total of 258 patients were enrolled between January 2016 and June 2023. Following the exclusion criteria, 226 patients (54.9% male) were included in the study. The process of participant enrollment is shown in Fig. 1, and the baseline characteristics are summarized in Table 1. The mean age was 73.2 ± 8.2 years, with 38.1% of patients aged older than 75 years. The mean CrCL was 53.7 ± 18.9 mL/min, with 44.7% of patients having CrCL < 50 mL/min. At the time of rivaroxaban concentration measurement, 86 patients (38.1%) were concurrently receiving CYP3A4 or P-gp inhibitors. These included amiodarone (n = 56), dronedarone (n = 15), diltiazem (n = 12), and verapamil (n = 6). Additionally, four patients (1.8%) were taking levetiracetam, a P-gp inducer. Furthermore, 22 patients (9.7%) were receiving BCRP inhibitors, specifically benzbromarone (n = 13), febuxostat (n = 8), and sulfasalazine (n = 1). The process of participant enrollment Table 1Demographic characteristics of the enrolled patientsCharacteristicAge (years)73.2 ± 8.2Male sex124 (54.9)Bodyweight (kg)65.0 ± 13.1Creatine clearance (mL/min)53.7 ± 18.9Laboratory testsa Blood urea nitrogen (mg/dL)21.5 ± 8.7 Serum creatinine (mg/dL)1.1 ± 0.3 Aspartate aminotransferase (U/L)23.9 ± 9.2 Alanine aminotransferase (U/L)21.4 ± 14.6 Total bilirubin (mg/dL)0.9 ± 0.6 Hematocrit (%)40.8 ± 4.9Comorbidities CHA2DS2-VASc scoreb3 (2, 5) HAS-BLED scorec2 (1, 2) Ischemic stroke/TIA history47 (20.8) Hypertension170 (75.2) Diabetes64 (28.3) Venous thromboembolism history5 (2.2) Gastrointestinal bleeding history8 (3.5)Rivaroxaban regimen and concentration 15 mg daily141 (62.4) 10 mg daily85 (37.6) Trough concentration (ng/mL)27.1 (15.6, 50.5) Time from last dose to trough sampling (h)24.6 ± 3.14 Peak concentration (ng/mL)232.8 (162.7, 343.2) Time after rivaroxaban administration to peak sampling (h)2.11 ± 0.40Concurrent medicationsd CYP3A4 or P-gp inhibitorse86 (38.1) Amiodarone56 (24.8) Dronedarone15 (6.6) Verapamil6 (2.7) Diltiazem12 (5.3) CYP3A4 or P-gp inducers (levetiracetam)f4 (1.8) BCRP inhibitors22 (9.7) Benzbromarone13 (5.8) Febuxostat8 (3.5) Sulfasalazine1 (0.4)Data are presented as mean ± standard deviation, a number (percentage), or median (interquartile range)BCRP breast cancer resistance protein, CYP cytochrome P, TIA transient ischemic attack, P-gp P-glycoproteinaProportion of missing data: Blood urea nitrogen (30.1%); serum creatinine (0%); aspartate aminotransferase (55.8%); alanine aminotransferase (28.3%); total bilirubin (82.7%); hematocrit (12.8%)bThe CHA2DS2-VASc score, which ranges from 0 to 9, was used to estimate the risk of thromboembolism during rivaroxaban therapy, with higher scores indicating a higher risk of thrombosiscThe HAS-BLED score, which ranges from 0 to 6, was used to estimate the risk of hemorrhage during rivaroxaban therapy, with higher scores indicating a greater risk of bleedingdDetailed classifications of concurrent medications are provided in Table S1eSome patients used more than one CYP3A4 or P-gp inhibitorfAll participants in this group received levetiracetam Demographic characteristics of the enrolled patients Data are presented as mean ± standard deviation, a number (percentage), or median (interquartile range) BCRP breast cancer resistance protein, CYP cytochrome P, TIA transient ischemic attack, P-gp P-glycoprotein aProportion of missing data: Blood urea nitrogen (30.1%); serum creatinine (0%); aspartate aminotransferase (55.8%); alanine aminotransferase (28.3%); total bilirubin (82.7%); hematocrit (12.8%) bThe CHA2DS2-VASc score, which ranges from 0 to 9, was used to estimate the risk of thromboembolism during rivaroxaban therapy, with higher scores indicating a higher risk of thrombosis cThe HAS-BLED score, which ranges from 0 to 6, was used to estimate the risk of hemorrhage during rivaroxaban therapy, with higher scores indicating a greater risk of bleeding dDetailed classifications of concurrent medications are provided in Table S1 eSome patients used more than one CYP3A4 or P-gp inhibitor fAll participants in this group received levetiracetam PPK Model Development A total of 12 blood samples exhibiting extreme rivaroxaban concentrations were excluded from the primary model development to prevent parameter distortion (Fig. 1). These included 11 samples with undetectable concentrations (10 trough and 1 peak) and 1 peak sample with an exceptionally high concentration of 1225.1 ng/mL, which exceeded five times the population median peak concentration. The observed concentration–time profile was adequately described by a one-compartment model with first-order elimination, which exhibited a mono-exponential decline without a discernible distribution phase (Fig. 2). The RV can be best described using a constant error model. IIV included the apparent volume of distribution (V/F) and CL/F, both of which were best described by an exponential model. This structure yielded the lowest OFV and satisfactory GOF plots (Fig. 3 and Table S3). Pharmacokinetic parameters of the base model are listed in Table 2. Observed rivaroxaban concentrations versus time after the last dose. The solid lines represent the linear regression mean for the 10 mg (blue) and 15 mg (orange) dose groups, with shaded areas indicating the 95% confidence intervals The diagnostic plots used for model evaluation. Base model: A Observed versus population-predicted concentrations. B Observed versus individually predicted concentrations. Final model: C Observed versus predicted population concentration. D Observed versus individually predicted concentrations. E Individual weighted residuals versus individual predicted concentrations. F Individual weighted residuals versus time since the last dose. In panels A–D, the solid identity line (y = x) represents a perfect prediction. In panels E and F, the horizontal line at zero indicates no prediction errors Table 2Pharmacokinetic parameters estimated from the base and final modelsBase modelFinal modelFixed effects Ka (h−1)0.82 (fixed)0.82 (fixed) V/F (L)41.83 (5.01)45.57 (6.21)LBW on V/F–0.01 (41.37) CL/F (L/h)5.34 (6.01)6.13 (6.21)CrCl on CL/F–0.64 (19.78)CYP3A4/P-gp inhibitors on CL/F–−0.38 (5.44)Interindividual variability ωV0.59 (6.15)0.59 (6.27) CV (%)64.8163.93 ωCL0.71 (7.22)0.59 (7.67) CV (%)81.4264.20Residual variability a17.50 (9.61)14.36 (17.72)Model fitting OFV4968.084914.67Data are presented as the estimated value (residual standard error)a residual unexplained variability (constant error model), CrCL creatinine clearance, CL/F apparent clearance, CV coefficient of variation, CYP cytochrome P, Ka absorption rate constant, LBW lean body weight, OFV objective function value, V/F apparent volume of distribution, ωV interindividual variability for V/F, ωCL interindividual variability for CL/F Pharmacokinetic parameters estimated from the base and final models Data are presented as the estimated value (residual standard error) a residual unexplained variability (constant error model), CrCL creatinine clearance, CL/F apparent clearance, CV coefficient of variation, CYP cytochrome P, Ka absorption rate constant, LBW lean body weight, OFV objective function value, V/F apparent volume of distribution, ωV interindividual variability for V/F, ωCL interindividual variability for CL/F Covariate Model Building The Apparent Volume of Distribution The covariate analyses identified body surface area (BSA), ideal body weight (IBW), LBW, LBM, and sex as potential factors significantly associated with V/F. V/F was larger in male patients and increased with higher BSA, IBW, LBW, and LBM. In forward inclusion, adding each of these covariates individually led to significant reductions in OFV, indicating improved model fit (Table S4). As BSA, IBW, LBW, and LBM were highly correlated, LBW resulted in the greatest reduction in OFV and was therefore selected for further modeling. During backward elimination, the removal of either LBW or sex did not meet the predefined level of significance (OFV changes of − 2.43 and − 0.97, respectively). Given the potential collinearity between LBW and sex, we first incorporated LBW alone and then excluded it to assess OFV changes, followed by incorporating sex alone and excluding it. The OFV changes were +8.66 for LBW and +10.10 for sex (Table S4). As both values exceeded the predefined threshold of significance (i.e., 6.64), both covariates were considered to contribute meaningfully to the model. Covariate selection for the final model was guided by GOF diagnostics. In the scatter plots of observed versus predicted rivaroxaban concentrations, the model incorporating LBW using a proportional-shift functional form demonstrated the best performance (R2 = 0.50); thus, LBW was retained in the final model. The Apparent Clearance The covariate analyses identified several potential factors that may influence CL/F, including age, IBW, LBM, LBW, CRE, eGFR, CrCL, sex, and concurrent use of CYP3A4/P-gp or BCRP inhibitors. CL/F was increased with higher bodyweight and better renal function (higher eGFR or CrCL), while it was decreased with aging or elevated CRE. In addition, CL/F was lower in female patients and those concurrently receiving CYP3A4/P-gp or BCRP inhibitors. During the forward inclusion process, the individual addition of the above covariates resulted in a significant improvement in the GOF (Table S5). Among IBW, LBM, and LBW, IBW was selected as the best body size descriptor on the basis of the greatest reduction in OFV. Similarly, CrCL was selected as a better predictor of the impact on renal function than CRE or eGFR. Accordingly, the full model incorporated age, IBW, CrCL, CYP3A4/P-gp inhibitor use, BCRP inhibitor use, and sex as covariates on top of the base model. During the subsequent backward elimination process, the removal of CrCL or the concomitant use of CYP3A4/P-gp inhibitors led to significant increases in the OFV (+6.75 and +16.48, respectively). This confirms that both covariates are essential components of the final model (Table S5). By comparing the GOF scatter plots of PRED versus the observed data, CrCL with a power function and the use of a CYP3A4/P-gp inhibitor with a proportional shift function achieved the best model performance with a coefficient of determination (R2) of 0.58. The estimated pharmacokinetic parameters are listed in Table 2. A visual inspection of the GOF plots is shown in Fig. 3. Both the base and final models demonstrated excellent agreement between the observed concentrations and IPRED, with R2 values of 0.999. The final model showed improved predictability on the basis of PRED, with a higher R2 than that of the base model (0.553 versus 0.470). The VPC results (Fig. S2) demonstrated that the final model reliably describes the observed data, with the vast majority of observed concentrations and their respective percentiles falling within the simulation-based 90% prediction intervals. This confirms the robustness of the model for clinical applications in the target population. Model Validation An independent cohort of 67 patients receiving rivaroxaban was enrolled to validate the established PPK model. One patient with an undetectable trough concentration and one with missing essential covariates (height) were excluded from the analysis. At the individual level, the predicted rivaroxaban concentrations correlated well with the observed values, with a mean absolute percentage error (MAPE) of 10.24%. However, at the population level, the predictive performance was suboptimal, with MAPE of 91.1%. Scatter plots of the predicted versus observed rivaroxaban concentrations are shown in Fig. S3. Exposure Response Analysis The AUC was calculated for each participant using an established PPK model. A total of 145 patients (64.2%) had estimated AUC values within the expected range reported in the clinical trials, 50 (22.1%) had estimated AUC values below the expected range, and 31 (13.7%) had estimated AUC values above the expected range. While estimated AUC distributions overlapped between patients receiving rivaroxaban 10 mg and 15 mg, those in the 15 mg group had a significantly higher average estimated AUC compared with the 10 mg group (3606.3 ± 2279.2 and 2976.5 ± 1920.3 ng × h/mL, P = 0.03, respectively, Fig. 4). The association between the selected clinical factors and the model-estimated area under the curve (AUC) of rivaroxaban. A and C Scatter plot illustrating the association between creatinine clearance (CrCL) (A) or lean bodyweight and estimated AUC (C), stratified by rivaroxaban dose (10 mg versus 15 mg). B and D Boxplots comparing estimated AUC values between patients with and without concurrent use of cytochrome P450 3A4 (CYP3A4) or P-glycoprotein (P-gp) inhibitors (B) and the gender effects (D), stratified by dose. In the scatter plots, orange and blue dots represent patients receiving 15 mg and 10 mg regimens, respectively. Solid lines indicate linear regression trends and shaded areas represent 95% confidence intervals We further investigated the association between estimated AUC and the covariates identified as significant determinants of CL/F and V/F, as shown in Fig. 4. A reduction in CrCL or LBW was associated with an increase in rivaroxaban estimated AUC, but the association between CrCL and estimated AUC was more pronounced than that of LBW. Additionally, patients concurrently using CYP3A4/P-gp inhibitors had significantly higher estimated AUC compared with those without such inhibitors, both in the 10 mg group (3670.8 ± 2205.9 versus 2387.7 ± 1414.9 ng × h/mL; P = 0.002) and in the 15 mg group (4498.0 ± 2721.9 versus 3160.6 ± 1885.0 ng × h/mL; P < 0.001). The difference in estimated AUC between was not statistically significant in the 10 mg group (male versus female: 2685.4 ± 1331.7 versus 3274.4 ± 2357.6 ng × h/mL; P = 0.16), whereas in the 15 mg group, female patients had significantly higher estimated AUC values than male patients (4272.5 ± 2753.5 versus 3112.9 ± 1708.8 ng × h/mL; P = 0.003). For all enrolled participants, we applied the established population of PPK model to estimate the drug concentrations at standardized timepoints: 3 h post-rivaroxaban administration for Cpk and 24 h post-administration for trough levels Ct. The average estimated Cpk was 274.48 ± 43.3 ng/mL and the average estimated Ct was 157.55 ± 47.27 ng/mL. Notably, the model-estimated Ct was higher than the observed trough concentration (Table 1). This discrepancy likely reflects the difference between the simulated 24-h trough and the actual trough concentrations measured at variable time intervals since the last dose in real-world clinical practice. When compared with the expected ranges reported in clinical trials, the distribution of estimated Cpk was as follows: 64 patients (28.3%) were below the range, 104 patients (46.0%) were within the range, and 58 patients (25.7%) were above the range. For estimated Ct, 28 patients (12.4%) were below the range, 186 patients (82.3%) were within the range, and 12 patients (5.3%) were above the range. During the observation period, nine SSE events and ten major bleeding events were recorded. After adjusting for clinical covariates, treating estimated Ct, Cpk, or AUC as continuous variables did not reveal significant association between exposure metrics and clinical outcomes (Table S6). When patients were further categorized according to expected ranges from clinical trials, the incidence rates across different estimated Ct, Cpk, and AUC groups are presented in Table S7, with corresponding Kaplan–Meier survival curves presented in Fig. 5 and S4. Patients with low estimated Ct values exhibited a trend toward a higher risk of SSE, with an incidence of 1.78 per 100 person-years compared with 1.01 in the within-range group (log-rank P = 0.66; HR and 95% CI 1.69 [0.35, 8.16], P = 0.51). Conversely, patients with high estimated Ct values showed a trend toward an increased risk of major bleeding, with an incidence of 2.88 per 100 person-years compared with 1.15 in the within-range group (log-rank P = 0.64; HR and 95% CI 2.49 [0.31, 20.18], P = 0.39). However, these differences were not statistically significant, even after adjusting for age, sex, CrCL, and CHA₂DS₂-VASc score for SSE, or HAS-BLED score for major bleeding. The adjusted HR (aHR) and 95% confidence interval (CI) for SSE and low Ct was 1.81 (0.37, 8.94; P = 0.65), and the aHR and 95% CI for major bleeding and high Ct was 2.73 (0.30, 24.71; P = 0.37). The association between estimated Cpk or AUC and clinical outcomes is less consistent and inconclusive. Kaplan–Meier survival curves for clinical outcomes. A Stroke or systemic thromboembolism (SSE) and B major bleeding across the different trough concentration groups. The green line represents patients with within-range Ct (trough concentration), blue line represents those with low concentrations, and the orange line represents those with high concentrations Sensitivity Analysis The performance of the PPK model, with or without the inclusion of extreme values and using Ka values of 0.617 h⁻1 or 0.821 h−1, is summarized in Table S8. Under the constant RV model, the GOF plots appeared similar across approaches; however, the lowest OFV was obtained when Ka was fixed at 0.821 and extreme values were excluded. The correlation between CL/F and different covariates is summarized in Table S9. Only BUN showed significant correlation. However, we have incorporated CrCL, which is a renal function related covariate in the final PPK model. We did not further test BUN in the final model. The evaluation of the relationship between CL/F and various potential covariates is summarized in Table S9. Among the laboratory parameters with missing values, only BUN demonstrated a statistically significant correlation with CL/F. However, since CrCL, a primary indicator of renal function, was already successfully incorporated into the final PPK model, BUN was not further evaluated as an additional covariate. Discussion This study established a PPK model using real-world rivaroxaban concentrations in patients with AF, and found that rivaroxaban exposure was influenced by renal function, concurrent medications, and lean bodyweight, with renal function showing the strongest effect. The developed PPK model is applicable in clinical practice for estimating rivaroxaban exposure. Additionally, both high and low rivaroxaban concentrations demonstrated a trend toward an increased risk of clinical events. Comparing the ROCKET-AF and J-ROCKET AF, our study population had a similar mean age (approximately 73 years). However, patients in our real-world cohort had lower CrCL (median 54 mL/min) than those in J-ROCKET AF (median 65 mL/min) and global ROCKET AF (median 67 mL/min). According to the PPK model developed in our study, the estimated apparent volume of distribution (V/F) was 45.96 L, which is lower than the values observed in Caucasian populations (79.7 L). The CL/F was 5.89 L/h, which is within the range (4.46–5.98 L/h) documented in a systematic review of 10 PPK studies conducted in Asian populations [21]. In covariate analysis, LBW was identified as a significant determinant of V/F. Increasing LBW from the median (48 kg) to the maximum (81.6 kg) led to a 47% increase in V/F, whereas reducing LBW to the minimum (24.9 kg) resulted in a 32% decrease. However, the influence of LBW on the AUC was less pronounced, as reflected by the relatively flat relationship between LBW and the AUC (Fig. 4). Interestingly, actual bodyweight did not significantly affect V/F, possibly because of the moderate-to-low tissue affinity of rivaroxaban, making LBW a more accurate predictor of distribution [4]. Although sex was also identified as a significant factor influencing V/F, it was excluded from the final model because of its high collinearity with LBW. Notably, among the patients receiving the 15 mg regimen, female patients exhibited significantly higher rivaroxaban exposure than male patients. This finding is likely attributable to the baseline characteristics of the female subgroup. Compared with male patients, female patients in the 15 mg group were older, had lower body weights, and had significantly lower CrCL. Given that CrCL and body size were identified as critical determinants of rivaroxaban pharmacokinetics in our model, these cumulative physiological differences likely drove the observed increase in exposure. Further investigations are warranted to determine whether this disparity is clinically significant regarding long-term safety and efficacy outcomes. Consistent with previous studies, CL/F was influenced by CrCL [6, 8, 21]. This is in line with the fact that approximately two-thirds of rivaroxaban is excreted renally [4]. An increase in CrCL from the median (54 mL/min) to the maximum (133 mL/min) resulted in a 75% increase in CL/F, while a decrease to the minimum (19.8 mL/min) led to a 46% reduction. CrCL also showed an association with the AUC. Patients with a CrCL between 30 and 50 mL/min and those with a CrCL below 30 mL/min had AUC 35.7% and 58.2% higher, respectively, than those with a CrCL above 50 mL/min. Concurrent use of CYP3A4/P-gp inhibitors reduced CL/F by 32% and increased AUC by 41.8%. Notably, these effects were primarily driven by amiodarone and dronedarone, which accounted for 82.6% of co-medication cases. A nationwide insurance database study in Taiwan reported that concurrent use of amiodarone and direct oral anticoagulants (DOACs) was associated with an increased risk of major bleeding [22, 23]. Our findings provide a pharmacokinetic basis for this clinical association. In the validation cohort, the model demonstrated good predictive performance at the individual level, whereas the population-level predictions were less accurate. This discrepancy may be attributed to the limited sample size and restrictions on patients with AF. The suboptimal MAPE observed in this study may reflect substantial interindividual variability or inherent differences between the development and validation cohorts, rather than limitations of the model itself. Nevertheless, the PPK model retained its clinical value. In real-world practice, standardized blood sampling at specific timepoints, such as peak or trough levels, is often impractical, particularly in acute care settings. This model allowed the simulation of standardized rivaroxaban concentrations, facilitating improved interpretation and comparison across individuals. With the simulation, approximately 80% of the trough concentrations in our study were within the expected range from clinical trials, whereas only 60% of the AUCs and 50% of the peak concentrations met expectations. This deviation may have occurred because only a single blood sample was collected during the absorption phase. Because absorption times can vary among individuals, a single measurement may not accurately capture the true peak concentrations. Notably, most drug concentrations that fell outside the expected range were lower than anticipated. This may be partly explained by the frequent use of off-label underdose regimens in real-world practice, driven by concerns over bleeding risks [24]. In our cohort, approximately 30% of the patients received off-label underdose regimens. As demonstrated in our previous investigation and in the current study, low rivaroxaban concentrations increased the risk of thromboembolic events [10]. Therefore, preemptive dose reduction in patients without a documented history of bleeding should be discouraged. Our study has several strengths, including the use of real-world data and the clinical validation of the model. However, this study had some limitations. First, the sample size was modest and limited to Asian patients with AF, which restricts generalizability. Second, owing to sparse sampling during the absorption phase, we were unable to estimate Ka directly from our dataset. Although the model demonstrated good fit using previously published Ka values, the peak concentrations were estimated at a standardized 3 h post-dose timepoint. Given that food intake can delay time to peak concentration and that fasting status was not documented in this real-world cohort, these estimates may not perfectly capture the true Cpk for all individuals. Third, genetic polymorphisms such as CYP3A4 or ABCB1 variants, which may contribute to pharmacokinetic variability, were not evaluated in this study. Fourth, model diagnostics revealed a high correlation between IPRED and observed concentrations (R2 = 0.999, Figs. 3 and S2), which may partially reflect epsilon shrinkage, a common phenomenon in PPK modeling with sparse sampling. While the PRED exhibited some systematic underprediction, the final model demonstrated an improvement in predictive power (R2 = 0.553) compared with the base model (R2 = 0.470), suggesting that the identified covariates successfully accounted for a portion of the interindividual variability. Lastly, the number of clinical events was limited, which, combined with the potential variability in Cpk assessment, may have precluded the detection of definitive statistically significant relationships between drug exposure and clinical outcomes. In conclusion, this study established a PPK model on the basis of real-world rivaroxaban concentrations in an Asian population, identifying CrCL, concomitant use of CYP3A4/P-gp inhibitors, and LBW as significant covariates influencing pharmacokinetics. Furthermore, this developed PPK model facilitates the estimation of individualized Ct from random samples. As patients with Ct outside the expected ranges demonstrated trends toward higher risks of SSE or major bleeding, the PPK model serves as a practical tool for standardized exposure assessment and the identification of patients at risk of treatment failure. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (PDF 952 KB) Acknowledgments We would like to thank National Taiwan University School of Pharmacy (NTUSP) Endowment Fund in support of the Platform for Clinical Mass Spectrometry and NMR Structure Elucidation and Chia-Chi Chang from NTUSP for the measurement of DOAC level. We would also like to acknowledge Yu-Ju Tseng and Liang Juan for their technical contributions to statistical analysis and refinement of the PPK model. Funding This work was supported by research grants from the National Science and Technology Council, Taiwan (112-2314-B-002-313) and National Taiwan University Hospital (114-S0112). Declarations Conflict of Interest Nian-Shiuan Lai, Chun-Jung Lin, Ching-Hua Kuo, Yu-Fong Peng, Sung-Chun Tang, Chih-Fen Huang, Shin-Yi Lin, and Shu-Wen Lin declare that they have no potential conflicts of interest that might be relevant to the contents of this manuscript. Author Contributions Nian-Shiuan Lai, Shin-Yi Lin, and Shu-Wen Lin contributed to the conceptualization, methodology, investigation, and formal analysis. Shin-Yi Lin drafted the original manuscript, and all the other authors critically reviewed and edited of the manuscript. Ching-Hua Kuo contributed to the resources for concentration analysis. Sung-Chun Tang and Chih-Feng Huang contributed to the conceptualization and methodology of the investigation. Shin-Yi Lin contributed to funding acquisition. Data Availability The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request. Ethical Approval The study protocol was approved by the International Ethics Committee of NTUH (No. 202201014RINB) and registered in the ClinicalTrials.gov database (NCT05333666). Consent to Participate All participants provided written informed consent prior to enrollment in the study. Consent for Publication Not applicable. Code Availability Not applicable. Contributor Information Shin-Yi Lin, Email: hsin924@ntuh.gov.tw. Shu-Wen Lin, Email: shuwenlin@ntu.edu.tw. References Associated Data Supplementary Materials Supplementary file1 (PDF 952 KB) Data Availability Statement The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.